Seminar Talks

"Understanding and interpreting image contents requires both recognizing
individual entities within an image and their relationships with each
other. While there has been some interest at visual relationship
detection in the past years, current models need fully labeled data sets
and are fully discriminative. Furthermore, recent work indicates that
"simple" feed-forward networks like convolutional neural networks are
not able to "understand" visual relations on a conceptual level. Rather,
it seems that they learn relationships by heart and can not generalize
them to novel scenes or objects. We propose using generative models to
facilitate learning visual relationships even in (partially) unlabeled
data sets. Additionally, we will incorporate attention mechanisms to
enable the system to focus on individual objects and relationships. With
the combination of generative models and attention mechanisms we hope to
avoid the drawbacks of "simple" feed-forward mechanisms which only get
to see the input once, while also being able to make use of unlabeled
images in learning visual relationships."

Related papers:
- Lu, Cewu, et al. "Visual relationship detection with language priors."
/European Conference on Computer Vision/. Springer, Cham, 2016.

- Santoro, Adam, et al. "A simple neural network module for relational
reasoning." /Advances in neural information processing systems/. 2017.

- Ricci, Matthew, Junkyung Kim, and Thomas Serre. "Not-So-CLEVR: Visual
Relations Strain Feedforward Neural Networks." /arXiv preprint
arXiv:1802.03390/ (2018).

Yong Wang: Towards Unsupervised Learning of Action-relevant
Representations Guided by Self-motion Information (advisor: Cornelius)

Erik Fließwasser: Transfer Learning of Face Representations Using a Semi-Supervised
Strategy (advisor: Pablo)

Ahmed Elshinawi: Multi-Attribute Face Generation Using a Conditional
BEGAN (advisor: Pablo)

Theresa Pekarek-Rosin: Effect of a Humanoid's Active Role during
Learning with Embodied Dialogue System (advisor: Matthias)

Joan Tomas Pape: Continuous Deep Reinforcement Learning
with Normalized Advantage Functions for Robot Grasping (advisor: Matthias)

Current automatic speech recognition systems make use of a single source of information about their input, viz. a preprocessed form of the acoustic speech signal, which encodes the time-frequency distribution of signal energy. And in a real environment, ambient noise and reverberation usually degrade quality and reliability of speech recognition. Articulatory features (AFs) composed of probabilities (or, more generally, scores) are abstract classes describing the most essential articulatory properties of speech sounds, e.g. voiced, nasal and rounded. They are independent of acoustic variation such as speaker-dependent spectral differences, background noise, and room reverberation.
In this talk, I will present and discuss how to extract AFs from speech using neural networks and apply AFs to robot speech recognition.

This thesis presents a new bio-inspired approach for robot motion. The idea is to model the natural movement generation of humans closely. The natural central pattern generators are modeled with two artificial neural networks. This is based on a medical approach predicting muscle activity of a walking human. Ideal muscle activation patterns for humans are known, for robots these need to be discovered. This is done using a genetic algorithm. Thus, not only the generation but also the learning is natural and adaptable.

Xiaomao: "Advances in neuroscience have long suggested that animal navigations are
related to place and head-direction cells which mainly inhabit in
animal's hippocampus and exhibit space-related firing properties. To
explain their functional roles in navigation, we propose a
self-organizing method for robot navigation based on learned place and
head-direction cells, which includes modeling hippocampal cells from
visual inputs, spatial learning through self-organization, neural
planning and action control. Preliminary results from simulation show
that our method can navigate a robot to the desired position smoothly
and efficiently. This spottalk will mainly elaborate how to implement
the developed navigation system on a real robot platform."

Burhan: "Fast online learning of continuous visuomotor skills in sparse reward
environment is an open challenge for existing reinforcement learning
algorithms. We address this by extending our Intrinsically motivated
Continuous Actor-Critic algorithm using representations learned with a
hierarchical SFA network and a proposed convolutional autoencoder-based
actor-critic network, where we obtain state-of-the-art results on the
task of reach-for-grasping from raw visual input."

Alexandra Lindt - Facial Expression Editing Using a Conditional Adversarial Autoencoder

Meike Zimmermann - Tide and Storm Surge Prediction for Hamburg with Artificial Neural Networks

Emil Gasanov - Haptic Object Classification with Convolutional Neural Network using humanoid Sensor Hand

The walking movement of humans is graceful and robust. This movement is the result of synchronization between the neural mechanisms which generate rhythmic motion, and the dynamics of the skeletal structure. The overall movement is optimized by high-level control centers. Drawing inspiration from this mechanism, a hierarchical controller for bipedal locomotion of robots is proposed in this thesis. Artificial central pattern generators (CPGs) mimic the behavior of the neural mechanisms which produce rhythmic motion in animals. Many existing methods use CPG networks for bipedal locomotion, but most of them focus solely on the CPGs. The proposed controller augments the functionality of a CPG network by adding a novel high-level controller on top of it. Thus, at the lower level, a CPG network with feedback pathways controls the individual joints. The parameters of the CPG network are optimized by a genetic algorithm. At the higher level, a neural network modulates the output of the CPG network in order to optimize the robot's movements with respect to an overall objective. In this case, the objective is to minimize the lateral deviation while walking. The neural network is trained using reinforcement learning. The proposed controller was successfully used to produce stable bipedal locomotion for the NICO robot in simulation. Results of experiments show that the high-level controller was able to improve the performance of the low-level CPG network. Additionally, by comparing the performance of CPG networks with and without feedback mechanisms, the relative effectiveness of low-level feedback has been shown. The proposed controller is not strongly coupled to a particular robot model and is modular in design. The results obtained, by using this controller in simulation, encourage its use on the physical robot in the future.

The thesis will fill the gap with existing research in emotion recognition and emotion modeling in social robotics by implementing the concept of personalization, and suggesting a stand alone architecture / model integrated with an already existing social assistant in a home environment. The re- search will focus on emotion recognition (primary and secondary emotions from text) and personalization in socially intelligent robots (having deep cognition models like humans and offer human style of social intelligence). With the concept of personalization come the challenges like learning, adap- tion and memory. Socially Assistive Robotics (SAR) systems should adapt to each users physical and cognitive abilities and in this process should learn the patterns from interactions with users (and other robots for example) to boost its own attempts at interaction with its environment; referred to as personalized social interaction [14]. The learning process for SAR should hence be adaptive. The SAR should be able to store the emotional input from user in some form of memory, and use that effectively when taking decisions. The proposed architecture would effect the software part as well as hardware of the digital social assistant e.g. through head gestures (nodding if happy and shaking if sad), through LEDs and lights etc. Such an integration of emotions recognition and modeling into the existing architecture of such digital social assistants can be used to analyze where emotions can be useful in the overall architecture, areas where the emotional data can be used in the architecture, as well as correlation of personalization and emotion recognition with the assistants performance experienced by its users.

Can robot nociception increase the human-likeness of grasping postures?

Abstract:

Today a multitude of different humanoid robots exist, built within university projects or by companies. A commonly used technique to perform tasks with a robot is to train a neural network, which handles the robot?s actuators to fulfill the task. However, in most cases, only the looks of the humanoid robot are human-like. During tasks robots have non-human-like postures and motions because the main focus is to fulfill the task in a fast and accurate way. Even if human-likeness is desired, there is no human-likeness evaluation tool, which is commonly used or accepted. This work compares the postures generated by different reinforcement learning conditions during a grasping task in regards to human-likeness, to determine if a condition involving nociception is able to generate more human-like postures than other conditions. For the human-likeness evaluation of robot postures, an ergonomic assessment is used under the hypothesis that humans tend to have grasping postures, that are not painful and do not strain muscles or joints more than needed.

Comparison of Design Methodologies of Echo State Networks

Abstract:

In this work, we compare the design methodologies of Echo State Networks proposed by Ozturk et al. (2007), Boedecker et al. (2009) and Yildiz et al. (2012). We test their capability to predict Mackey-Glass time chaotic series and to learn Elman's grammar, which are standard benchmarks.

Bio-Inspired Auditory Signal Processing for Speech Recognition

Abstract:

With the advances in technology, speech is becoming a standard way to interplay with computers. For the human being, communication through speech has been around for thousands of years. Years of evolution have reproduced a fascinating speech recognition system that performs well in noisy environments, while Automatic Speech Recognition tools perform poorly. However, it is an inevitable problem for computers to endure noisy environment. Known models of the peripheral auditory system and speech enhancement neural approaches have the potential of improving verbal interaction with a computer. This thesis presents the possibility of improving the Word Error Rate of the Google Speech Recognition web API with the introduction of a bio-inspired filter.

In the task of text classification, models like Word2Vec and Doc2Vec can be used to produce and shape a vector representation of words and documents. Such vectors can be used in a classifier, like a multilayer perceptron, to automatically classify new texts into different categories. For this work, we are working with chat dialogs, which are special forms of text. Their structure differs from commonly used texts for such a task. Chat dialogs are shorter, contain spelling errors, and are written by two authors. This work will find out how to use Word2Vec and Doc2Vec as a base for a new model that produces vector representations for chat dialogs and is used successfully in an MLP. Therefore different models were tested with different parameter settings. The best models then were tested with the classifier. We have found two models that yield to an accuracy of around 90% for the MLP. The first model uses the word vectors trained by Word2Vec to produce a vector for chat dialogs. The second model combines two Doc2Vec models and merges the dialog vectors into one.

Instance-based object detection and fine pose estimation is still an active research problem in computer vision. While the traditional interest-point-based approaches perform reasonably well under controlled environments and for rigid objects with detailed textures, those approaches fall short when the environment is less-constrained or the objects of interest are texture-less or partially-deformable. On the other hand, CNN-based approaches have shown impressive results in more general object recognition tasks like image classification and category-based object detection and coarse pose estimation, with strong generalization abilities across different object properties and scenes. Still, their performance in those general problems is not quite at the stage of very high precision. This thesis looks into the possibility of getting the best of both worlds, by reducing the scope of application of CNNs from category-based to instance-based while in the same time maximizing their performance and achieving instance-based fine pose estimation in situations where traditional interest-point approaches perform poorly. The main problem of applying a CNN-based approach in this scenario is that it requires a lot of precisely annotated training data and producing such data manually requires a lot of effort. To overcome this problem, a typical approach is to simulate the whole image acquisition process and produce a large amount of training data. But the question that arises is whether the model trained on such synthetic data is applicable in real-world scenario. To analyse this question, we propose a CNN-based pose estimation model that is trained with synthetic data produced from the (possibly imperfect) 3D models of the objects. We also propose a methodology under which we can carefully examine how different parameters of the model, the synthetic data and the training procedure influence on the results of the model on real-world data and whether the model trained with synthetic data can be successfully applied on actual images of the objects. Our results show that the proposed model can be trained with synthetic images of the objects and still be successfully applied in a real-world scenario. Based on the results, we present more general insights about training neural models with synthetic images for application on real-world images.

Inspired by how the human brain applies saccadic masking to suspend the processing of blurred visuals during movement of the eyes, this thesis lays the foundation for a selective approach of delaying the capturing of camera frames on a humanoid robot in order to avoid phases in the gait that produce heavily blurred images. Beneficial retarding of individual frames in contrast to skipping them increases the average sharpness of images taken, without significantly decreasing the frame rate. Being a passive technique, the procedure is especially suitable for simple robots which otherwise lack capabilities like mechanically stabilising the imaging device or applying reconstructive post-processing. In order to quantify the ability of the platform to capture sharp images at a particular moment, we neurally predict the sharpness of frames from the current and prior values of the inertial and joint sensors and the actuators of the robot. We use a no-reference (NR) metric that produces comparable values across images with varying content to extract sharpness values from captured images. The results are then used to train an echo state network (ESN) with different learning techniques. The sensor values used as inputs are sampled at a higher rate than the camera frames from which the target values for the neural training are derived. Hence, the training data contains inputs for which the correct outputs are unknown. In this thesis, we develop an offline learning algorithm that copes with such missing target values. We compare it to online learning techniques trained on incomplete data and show that it performs comparable to training on complete data sets. The amount of images that can be captured continuously is limited by the memory of the robot. This leads to short sequences of training data. We show that looping these sequences enables comparable performance to learning in conventional setups. Forecasting image sharpness from the sensor values of the robot proved to be challenging. Due to the robot drifting while walking, the motives of the recorded images varied widely, from clear views of the surroundings to featureless close-ups of obstacles. This created a complex target signal for the training. The precision of the sharpness values forecast by the network trained in this thesis are insufficient for it to be deployed on a robot. The developed offline learning algorithm, however, is able to reproduce recognisable features of the target function.

The brain integrates information from multiple sensory modalities yielding a coherent, robust, and efficient interaction with the crossmodal environment. Multisensory interaction is progressively fine-tuned throughout the lifespan. Low-level stimulus characteristics (e.g., spatial proximity and temporal coincidence) are available before the formation of learned perceptual representations to bind increasingly more complex higher-level characteristics (e.g. semantic congruency). From a computational perspective, multisensory integration from spatio-temporal properties of stimuli has been widely studied, resulting in a generous number of subcortical processing approaches such as the superior colliculus. Similarly, approaches for binding high-level representations at a cortical level have been proposed. However, cortical and subcortical models have been proposed separately and the interaction of cortico-collicular representations has received considerably less attention.

In this talk, I will present and discuss a model developed in collaboration with the Biological Psychology and Neuropsychology group of UniHH. We propose a neural network architecture that learns the interaction of spatial-associative representations to perform causal inference and binding of audiovisual stimuli. This model shows that complex multisensory functions can emerge as a property of self-organizing learning in a cortico-collicular architecture. In the subcortical layer, topographically arranged networks account for the interaction of audiovisual stimuli in terms of population codes. In the cortical layer, congruent audiovisual representations are obtained via the experience-driven development of feature-based associations. Our novel proposal is that activity-driven levels of associative congruency (obtained as a by-product of the self-organizing neurodynamics) can be used as top-down modulatory projections to the subcortical layer. This results in semantically related audiovisual pairs yielding a higher level of integration than unrelated pairs (unity effect). I will present current results and a series of planned experiments to validate our model for multisensory interaction in complex environments.

As the desire for a flawless human-robot interaction arises from the accelerated technological progress, robots need a sophisticated understanding of natural language when interpreting the instructions given by a person. However, due to the fact that natural language is ambiguous by design, and human speech is transmitted via a noisy channel, issues can always emerge. Therefore, an internal representation of the interpretation of an instruction needs to be flexible, but still appropriate for evaluating the exact parameters for executing it. In this thesis a declarative representation, which uses logical predicates to express relations, is evaluated in terms of feasibility for inferring execution parameters of an instruction, and in terms of perceptibility of ambiguity. This is realised in the context of a 3D Blocks World, for which collected commands directed at a robotic arm were annotated with semantic interpretation data. Each command directed the robot to move a single object out of a constellation of cubes and prisms with different colors. The conducted experiment, consisting of the evaluation of the presumed start and goal position of the moved object for all commands, shows that the chosen representation is suited for the aforementioned task. The implemented system was able to immediately find the correct solution in 99.67% of the commands in the test data, while being able to spot ambiguous input in the remaining cases. Additionally, a concept for disambiguating the remaining commands is formulated.

Fares Abawi - Surprisal in Recurrent Neural Networks (Independent study, advisor Tayfun)

Sebastian Springenberg - Attention Mechanisms in Neural Encoder-Decoder Architectures (Independent Study, advisor Cornelius)

Nikhil Churamani - Modelling Affective Cores for Action Modulation in Social Robots (Independent Study, advisor Pablo)

Franziska Papenhausen - Artificial neural networks for collision detection and current intensity prediction in robotic motors (BSc thesis, advisor Sven)

Tobias Knöppler - Hierarchical Object tracking (HIOB) on the NICO robot (BSc thesis, advisor Stefan H.)

Patrick Eickhoff - Speech Enhancement Generative Adversarial Network for the WTM Robots (BSc thesis, advisor Cornelius)

Cognitive control (also called “executive function “) is a form of top-down information processing with functions like storage, plan and manipulation to successfully produce the appropriate response depending on the current goals or long-term plans. When people have to overcome automatic behavior (like reward seeking, addiction, etc.), or the current task is unfamiliar, dangerous, or related with decision-making processing, cognitive control plays a role of vital importance. In psychological experiments, researchers explore the inner mechanisms of cognitive control by studying cognitive conflict processing. Stimulus-response compatibility (SRC) tasks are always adopted and modified by controlling parameters like conflict types or conflict effect to mimic complex conflict circumstances in the real life. My talk will present previous work and findings in the field of cognitive conflict processing, including the paradigms we have designed, developing behavioral patterns, task-related fMRI and resting-state fMRI results, and theory models we had proposed. Finally, I will introduce the plan and progress of some new A5 experiments in the context of the CML project.

This talk presents the inverse kinematics solution using neural networks for a robotic arm in 3-dimension. The presented algorithm uses a pre-trained neural network that has been trained beforehand to solve forward kinematics to fine-tune another neural network which is learning inverse kinematics. Therefore, a neural network is firstly trained to learn the forward kinematics. So with a given sequence of joint configurations the trained neural network is able to return a position and orientation in Cartesian space. Afterwards, the pre-trained neural network will be fully connected with another feed-forward multilayer perceptron. This means, the input layer of the pre-trained network is connected with the output layer of the network that is to be trained. Finally, the new neural network, consisting of an untrained structure and a trained one (pre-trained neural network) in the rear part, can be trained supervised using pairs with identical position and orientation. Due to this architecture, the neural network in the front (untrained) will be trained through back-propagation to solve inverse kinematics. The algorithm suggests a transferable concept for other problems in the same category as inverse kinematics. Furthermore, this algorithm has the advantage of using unprepared training samples. The results show that this approach leads to good accuracy for the tackled learning problem.

Bedridden patients can present physical and mental illnes because of this condition. There are some protocols for avoiding such issues, but either they are to expensive, sensitive to human error or just not applied because of its complexity. We propose the development of a system with techniques of artificial intelligence and pattern recognition to identify the emotional state and posture of bedridden patients, based on the identification of facial and body gestures, called SIAPA (Sistema Inteligente de Apoio ao Paciente Acamado, Intelligent Patient Support System ). The recognition of emotion will enable the patient's emotional state to be inferred so that the proposed system can interact with it and recommend activities that combat a possible depression. Recognition of body gestures will allow SIAPA to identify the patient's posture on the bed over time, verifying the possibility of the appearance of pressure ulcers due to the fact that the patient remains in the same position for a long period of time. In this presentation, we will show the developmental steps of the SIAPA and give a brief introduction about the University of Pernambuco, host of such project.

Relation extraction has been widely used for finding unknown relational facts from the plain text. Most existing methods focus on exploiting mono-lingual data for relation extraction, ignoring massive information from the texts in various languages. To address this issue, we introduce a multi-lingual neural relation extraction framework, which employs mono-lingual attention to utilize the information within mono-lingual texts and further proposes cross-lingual attention to consider the information consistency and complementarity among cross-lingual texts. Experimental results on real-world datasets show that our model can take advantage of multi-lingual texts and consistently achieve significant improvements on relation extraction as compared with baselines.

Advancements in Human-Robot Interaction involve robots being more responsive and adaptive to the human user they are interacting with. For example, robots model a personalised dialogue with humans, adapting the conversation to accommodate the user's preferences in order to allow natural interactions. This study investigates the impact of such personalised interaction capabilities of a human companion robot on its social acceptance, perceived intelligence and likeability in a human-robot interaction scenario. In order to measure this impact, the study makes use of an object learning scenario where the user teaches different objects to the robot using natural language. An interaction module is built on top of the learning scenario which engages the user in a personalised conversation before teaching the robot to recognise different objects. The two systems, i.e. with and without the interaction module, are compared with respect to how different users rate the robot on its intelligence and sociability. Although the system equipped with personalised interaction capabilities is rated lower on social acceptance, it is perceived as more intelligent and likeable by the users.

Humans are exceptionally good at understanding and effectively describing complex visual scenes, i.e. to infer the salient entities and their characteristics. They can deduce how objects in the scene interact and what kind of relationship they have with each other. Humans are able to learn this in a mostly unsupervised manner and can even identify and focus on certain aspects of a scene based on external contexts such as questions. Capabilities like this are needed to understand complex visual scenes and to extract knowledge from them, e.g. to answer questions or to reason about a scene. In order to acquire these skills we need to develop methodologies that can learn the underlying concepts and relationships, preferably in an unsupervised way. For this we need good representations for visual scenes which enable us to reason about the underlying concepts. We propose a framework based on generative models, that can learn interpretable and modifiable representations in a mostly unsupervised manner based on scenes and associated data, e.g. questions or commands. These representations then enable us to reason about individual concepts and allow for modifying these concepts and contents of a scene in a controlled manner. Finally, the methodologies developed for learning these representations can then be used for a multitude of applications, such as un-/semisupervised learning, scene labeling (conditioned on external information), command execution, visual question answering, data augmentation, and data generation.

This talk introduces the techniques used in an interaction scenario, realised using the Neuro-Inspired Companion (NICO) robot. NICO engages the users in a personalised conversation where the robot always tracks the users' face, remembers them and interacts with them using natural language. NICO can also learn to perform tasks such as remembering and recalling objects and thus can assist users in their daily chores. The interaction system helps the users to interact as naturally as possible with the robot, enriching their experience with the robot, making it more interesting and engaging.

Josip Josifovski - Object detection and pose estimation with convolutional neural networks learned from synthetic data (advisor Matthias)

Maham Tanveer - Personalization in a digital social assistant (advisor Sascha)

Andreas Grenzing - Neuro-Symbolic Learning of Action Planning (advisors Matthias, German)

Human-Robot Interaction (HRI) has been receiving increasing attention from scientific communities and industries. This attention can be explained by the growing number of robots in our daily lives, from robots that share workspace with humans in the industrial sector to sociable companion robots living with the elderly in nursing homes. In the context of social robots, one of the most difficult challenges is to make the communication capability of robots as natural as possible. Although the modeling of emotional behaviors through emotion recognition demonstrated several benefits and significant improvements in HRI, the majority of recent approaches have used strongly supervised learning techniques that are not able to adapt to the large variation between individuals and situations in real world applications. A robot capable of understanding these variations could conduct a more natural interaction by given the best-tailored response to a person, specially to the elderly in which the emotional variations are less evident due to physical or cognitive fatigue. To give a robot not just the ability to perceive emotions, but also to adapt and respond accordingly to a person and a situation, we propose an adaptive emotional framework for social robots that will be able to learn emotional variations through interactions using face and upper-body features. The framework will be composed of different artificial neural architectures, including a feedforward for face detection and emotion recognition, an auto-associative for face recognition, and a recurrent architecture for decision-making. The neural networks will be pre-trained using public datasets, and their training will be continued during social interaction. Finally, the proposed framework will be evaluated using the Interaction Quality concept that is the definition of a single metric to measure the effectiveness of an HRI.

Emotion is a vital aspect of establishing empathy and understanding between humans in interactions, which in turn leads to humans experiencing those interactions more favourably. In order for robotic agents to fit more naturally into the daily life of humans, they should possess an ability to detect emotions in humans and respond to these emotions accordingly. The goal of responding to these emotions is to reduce frustration and improve interaction quality in Human-Robot Interaction (HRI). Speech is one of the primary forms of communication between humans and is also one of the most emotionally salient, making it particularly interesting when attempting to classify emotions in humans. We will present an overview of our research, which focuses on finding emotional features in speech, primarily in language, and how we intend on using an ensemble architecture to analyse these features from temporal and symbolic perspectives. Furthermore we will touch on the subject of how, once found, emotions can be used to adapt and steer dialogue to make HRI interactions more enjoyable for humans.

Der NICO-Roboter des Arbeitsbereiches Knowledge Technology ist ein humanoider Roboter, der für diverse Aufgaben entwickelt wurde. Durch Bewegungen seiner Gliedmaßen können Kollisionen mit Hindernissen entstehen. Der Roboter besitzt jedoch keine Kollisionserkennung, um reagieren zu können. Dies kann dazu führen, dass Schäden am Roboter entstehen. Diese Problematik diente dieser Arbeit als Motivation. Kommt es zu einer Kollision, steigt die Stromstärke der Motoren in den Gelenken des Roboters an. Mittels Vorhersagen der Stromstärke, können Rückschlüße auf eine mögliche Kollision getroffen werden. Zunächst wurde ein MLP trainiert, welches die Stromstärke eines Kopfmotors vorhersagen konnte. Anschließend wurde ein weiteres MLP trainiert, mit dessen Hilfe eine Vorhersage von vier Motoren im Arm getroffen werden konnte. Beide Netze lieferten vielversprechende Ergebnisse. Die Kollisionserkennung wurde mittels der Vorhersagen und einem weiteren MLP durchgeführt. Diese erkannte erfolgreich Kollisionen von bestimmten Bewegungsabläufen.

We investigate learning approaches, more specifically interactive reinforcement learning to perform a domestic task. We use parent-like advice to explore two set-ups: agent-agent and human-agent interaction. This thesis contributes to knowledge in terms of studying the interplay of multi-modal interactive feedback and contextual affordances. Overall, we investigate which parameters influence the interactive reinforcement learning process and show that the apprenticeship of reinforcement learning agents can be sped up by means of interactive parent-like advice, multi-modal feedback, and affordances-driven environmental models.

Mnemosyne is a team in computational neuroscience in Bordeaux, hosted in the NeuroCampus with neuroscientists and medical people. Our scientific positioning is about Mnemonic Synergy (hence Mnemosyne), the study of interactions between different kinds of learning and their organization in the brain in a system of memories. I will explain in more details this positioning and give some examples of current researches in the team, including modeling pavlovian conditioning by considering the different kinds of afferences of the amygdala and modeling operant conditioning through interactions between the prefrontal cortex and the basal ganglia.

Yang Li - Comparison of Design Methodologies of Echo State Networks

José Rodriguez - Bio-inspired Auditory Preprocessing for Speech Recognition

Tim Reipschläger - Learning Inverse Kinematics with Human-like Poses

Fabio Wendt - Vector Representation of chat dialogues for text categorization

Julian Tobergte - Improved Interpretation of Robot Commands via Crossmodal Validation with a Spatial Planner

Spoken language is one of the most efficient ways to instruct robots about performing domestic tasks. However, the state of the environment has to be considered to successfully plan and execute the actions. We propose a system which can learn to recognize the user's intention and map it to a goal for a reinforcement learning (RL) system. This system is then used to generate a sequence of actions toward this goal considering the state of the environment. Symbolic representations are used for both input and output of a Deep RL module. To show the effectiveness of our approach, the TellMeDave corpus is used to train the intention detection model and in a second step train the RL module towards the detected objective represented by a set of state predicates. We show that the system can successfully recognize instructions from this corpus and map them to the corresponding objective as well as train an RL system with symbolic input.

Present approaches to ASR consist of two main models: the acoustic and language model, which are being jointly searched for an optimum solution. Where acoustic models are largely task independent, language models can be restricted to particular tasks. The hybrid ASR-system DOCKS combines strong acoustic models provided by Google Search by Voice and the stationary ASR-system Sphinx-4 to improve the result provided by Google with domain-specific language models and restricted vocabularies. DOCKS can be operated with high ordered statistical n-gram models, but it is currently limited to a flat representation of the search space. This leads to extensive creation of search states and, therefore, to computational inefficiency during decoding, which causes the approach to perform slow regarding execution times. This thesis presents a lexical tree dictionary representation for the search space, which reduces the number of search states and the complexity during decoding of DOCKS and empowers DOCKS to run in real-time. Our experiments show that this approach outperforms the flat representation in execution times for large restricted vocabularies. The results also show that the performance in terms of word error rate is not negatively affected.

As the efforts to make human-robot interaction more sophisticated are progressing, large data sets are needed to improve robots’ language capabilities. These data sets need to be of high diversity to make a robot understand a great variety of phrases. Appropriate data is usually specific for a task, and cannot be found in existing data sets, so it has to be collected. Research teams usually do not manage to find a large and diverse crowd to collect the data, so the data sets end up to be monotonous, and robots can only learn a small variety of sentences from them. In recent years, crowdsourcing has emerged as a method to collect large data sets from all over the world online. However, the quality of the generated data is poor, and the workload required to evaluate it too high to make crowdsourcing worthwhile. This thesis shows that crowdsourced data quality can be improved through tutorials and assessment of the work of others, providing simple techniques to make crowdsourcing more feasible for the collection of data sets. The findings suggests that crowdsourcing could establish itself as a method of data collection in the field of human-robot interaction, opening many new possibilities for the lingual training of robots.

Tayfun Alpay: Overview of Current Framework Landscape

Pablo Barros: Feedforward Neural Networks in Keras

Egor Lakomkin: Recurrent Neural Networks in Keras

Erik Strahl: Running Keras, Theano and Tensorflow on Knowledge Technology Servers and Special Workstations - Basic Concept and Usage

Manfred Eppe: A Guide to Fast and System-Independent Deployment and Development of GPU-Enabled Tools with Docker

In a conversation, humans use changes in a dialogue to predict safety-critical situations and use them to react accordingly. We propose to use the same cues for safer human-robot interaction for early verbal detection of dangerous situations. The sentiment helps to guide a decision-making process in the verbal interaction, which can be used as a feedback signal in the dialogue-based context learning. Due to the limited availability of annotated dialogue corpora, we use a classification tool for utterances to neurally learn sentiment changes within dialogues and ultimately predict that of the upcoming utterances. We train a recurrent neural network on context sequences of words, defined as two utterances of each speaker, to predict the sentiment class of the next utterance. Our results show that the model learned safety-cues which can be used for a safer human-robot interaction.

Object tracking in video streams is a requirement in many fields of computer science. It is a challenging task because it requires an extraction of visual features from the video stream to identify the tracked object, as well as the ability to adapt to a changing appearance of the object over time. Convolutional Neural Networks (CNNs) have been proven to be powerful at extracting object representing features in image data, and, being a Machine Learning method, they are also able to adapt to a changing of those features over the tracking's duration. The thesis provides a thorough analysis of how a solid CNN based object model can be created during the tracking process by obtaining training data from the video stream. An update strategy is developed that incrementally trains the central CNN time efficiently and avoids bad data samples from being used. A deeper understanding of how the CNN is able to provide an adaptive model is gained throughout the development. In order to conduct the necessary research, a modular tracking framework is constructed, that achieves state of the art results in two commonly used tracking benchmarks.

Various ways of robotic based package delivery are currently developed in ifferent companies, in start-ups as well as in major enterprises. However, these systems do not collaborate with humans to perform their delivery. In collaborative human-robot tasks, it is of utmost importance that the robot performs the right action at the right time. This behaviour control is typically implemented using the Belief-Desire-Intention-concept (BDI) or state machines. Unfortunately, there is no well accepted common guideline when to use which one of these two approaches. From the literature it follows that state machines are often used for smaller, manageable applications and the BDI concept for hierarchically nested problems. In this thesis, we investigate a collaborative package delivery scenario and evaluate if state machines or the BDI-concept are better suited for this use case. The behaviour, as central module of the use case, has been implemented with PROFETA, a BDI-framework, and SMACH, a state machine implementation for ROS. Both implementations have been analysed and compared in different categories. Our results show, that the BDI-framework was easier to implement, it needed less characters than the state machine implementation. In the dynamic comparison, where the CPU and memory usage have been analysed, both frameworks performed similarly. A big difference has been detected in the comparison of subjective parameters. SMACH is easier to use because tutorials are provided and it is supported by a community. Overall, SMACH should be preferred as framework for the implementation of small use cases like the one presented in this thesis. The BDI concept should be used for larger applications, with an implementation that provides better subjective values than PROFETA.

A central goal of Human-Robot Interaction (HRI) is to enable robots to be part of our everyday lives. As gestures play a big role in human communication, gesture recognition is an essential task robots should be capable of. Existing gesture recognition work mostly focuses on scenarios with one person in the scene. However, scenes with more than one person are of particular interest, as robots may face multiple people in natural environments. Motivated by the ability of humans to focus on important parts of a scene, this thesis proposes an attention mechanism for dynamic body gestures. For this task, we designed two neural architectures with 2D and 3D convolution kernels. Additionally, we present an approach to train the proposed architectures using a generated dataset. We created this dataset by combining recordings of gestures and generated artificial scenes with multiple people. Our experiments showed that the proposed networks are able to learn an attention mechanism for gestures. 2D convolutions performed well for body gestures with distinct arm positions. 3D convolutions use spatio-temporal features which showed an advantage on subtle gestures with less characteristic arm positions. To demonstrate a potential application of our architectures, we present a concept study to apply the proposed attention mechanism to the task of gesture recognition.

We investigate Echo State Networks (ESN), which implement a new training paradigm for recurrent neural networks. We first demonstrate their gesture classification performance considering two feature sets with very distinct complexity. Second, we introduce the recurrence analysis for qualitative and quantitative description of the gesture input and the system dynamics of an ESN, and show that the methodology complements classic stability measures. Finally, we address the reservoir itself and propose an algorithm for pruning connectivity in a one-shot learning scenario.

State-of-the-art speech recognition systems steadily increase their performance using different variants of deep neural networks and postprocess the results by employing N-gram statistical models trained on a large amount of data coming from the general-purpose domain. While achieving an excellent per- formance regarding Word Error Rate (17.343% on our Human- Robot Interaction data set), state-of-the-art systems generate hypotheses that are grammatically incorrect in 57.316% of the cases. Moreover, if employed in a restricted domain (e.g. Human- Robot Interaction), around 50% of the hypotheses contain out-of- domain words. The latter are confused with similarly pronounced in-domain words and cannot be interpreted by a domain-specific inference system. The state-of-the-art speech recognition systems lack a mecha- nism that addresses syntactic correctness of hypotheses, which would be somehow comparable to the P600 Event-Related- Potential (ERP) that could be recorded from the brain. We propose a system that is inspired by this P600 ERP and can detect and repair grammatically incorrect or infrequent sentence forms. It is inspired by a computational neuroscience model that we developed previously. The current system is still a proof- of-concept version of a future neurobiologically more plausible neural network model. Hence, the resulting system postpro- cesses sentence hypotheses of state-of-the-art speech recognition systems, producing in-domain words in 100% of the cases, syntactically and grammatically correct hypotheses in 90.319% of the cases. Moreover, it reduces the Word Error Rate to 11.038%.

Emotions are related to many different parts of our lives: from the perception of the environment around us to different learning processes and natural communication. Therefore, it is very hard to achieve an automatic emotion recognition system which is adaptable enough to be used in real-world scenarios. This paper proposes the use of a growing and self-organizing affective memory architecture to improve the adaptability of the Cross-channel Convolution Neural Network emotion recognition model. The architecture we propose, besides being adaptable to new subjects and scenarios also presents means to perceive and model human behavior in an unsupervised fashion enabling it to deal with never seen emotion expressions. We demonstrate in our experiments that the proposed model is competitive compared with the state-of-the-art approach, and how it can be used in different affective behavior analysis scenarios.

A central goal of Human-Robot Interaction (HRI) is to enable robots to be part of our everyday lives. As gestures play a big role in human communication, gesture recognition is an essential task robots should be capable of. Existing gesture recognition work mostly focuses on scenarios with one person in the scene. However, scenes with more than one person are of particular interest, as robots may face multiple people in natural environments. Motivated by the ability of humans to focus on important parts of a scene, this thesis proposes an attention mechanism for dynamic body gestures. For this task, we designed two neural architectures with 2D and 3D convolution kernels. Additionally, we present an approach to train the proposed architectures using a generated dataset. We created this dataset by combining recordings of gestures and generated artificial scenes with multiple people. Our experiments showed that the proposed networks are able to learn an attention mechanism for gestures. 2D convolutions performed well for body gestures with distinct arm positions. 3D convolutions use spatio-temporal features which showed an advantage on subtle gestures with less characteristic arm positions. To demonstrate a potential application of our architectures, we present a concept study to apply the proposed attention mechanism to the task of gesture recognition.

Understanding other people's actions plays a crucial role in our everyday lives. Human beings are able to reliably discriminate a series of socially relevant cues from body motion, with this ability being supported by highly skilled visual perception and other modalities. The main goal of this thesis is the modeling of artificial learning architectures for action perception with focus on the development of multimodal action representations. As a modeling foundation to address our research question, we focus on hierarchies of self-organizing neural networks motivated by experience-driven cortical organization.

Deep learning is dependent on large amounts of annotated training data. For the development of robotic visuomotor skills in complex environments, generating suitable training data is challenging and time-consuming. Deep reinforcement learning tries to alleviate this problem by letting the robot learn unsupervised through trial and error; this, however, comes at the cost of long training times. In this talk, an approach for learning visuomotor skills from explorative interaction with the environment is presented. The robot generates its training data based on initial motor skills acquired through demonstration. End-to-end learning of visuomotor skills is realized with a deep convolutional neural network that combines two important subtasks of grasping, object localization and inverse kinematic, in one integrated architecture.

Robot navigation is the core technology and also a difficult task related to robot perception, planning, implementation and many other aspects in mobile robot research. Different from probabilistic approaches which are mainly based on Kalman or Particle Filters, bio-inspired robot navigation is inspired by navigation behavior in animals and makes hypotheses about how the navigation skills are acquired and implemented in an animal's brain and body. Thus it is a promising direction to mimic animal behavior and attempt to achieve their learning ability for intelligent robots. Head-direction cells, place cells and grid cells have been found to play a fundamental role in rat navigation and have been computationally modeled. we aim to develop a navigation system which is based on these biological findings for a real Jackal robot platform. Essentially, robot navigation requires the capabilities of localization, map building, path planning and obstacle avoidance. We intend to just use robot's own cameras and odometry to complete these work, which can significantly simplify the system.

Reinforcement Learning (RL) helps robots learn behaviors including complex manipulation and navigation tasks autonomously, purely from trial-and-error interactions with the environment and resulting in an optimal policy, which represents the desired behavior of the robot. However, exploring the environment to collect useful experience data is often difficult, especially for grasping tasks where high-dimensional visual data and multiple degrees of freedom are involved. The PhD project proposes to address this RL problem in continuous state and action space by looking into exploration strategies inspired by infant cognition and sensorimotor development, particularly the use of curiosity as an intrinsic motivation. Curiosity plays an important role in child development and we believe it has similar impact on robot learning of grasping skills.

For cognitive science robotics is an interesting tool for studying the human mind while at the same time working on solutions to real world applications of state-of-the-art research. Both social robotics and the related field of developmental robotics have become highly interdisciplinary and one of the main challenges here is that this type of research needs to be informed by both an understanding of the social science and the engineering component. To achieve more natural interaction with robotic systems, inspiration may be sought in the more advanced field of intelligent virtual agents (IVA) or embodied conversational agents (ECA). This community is in closer interaction with the dialogue systems community than the human-robot interaction scene. Further, these interfaces are much more developed with respect to speech-gesture alignment, social variation and naturalness. The main difference, however, is that whereas the robotics community often treats user interfaces and especially speech-interfaces as add-ons in most architectures, the interaction is at the heart of ECAs and tightly integrated into the software architecture. What needs to be achieved for similar systems in robotics is a transfer from the virtual to the physical, taking "real world" constraints of the physical embodiment into consideration. Finally, human-robot interaction can benefit from good research and experiment design in the style of psychological studies.

In recent advances a new type of semi-supervised neural networks, named Ladder Networks, made it possible for neural networks to utilize unlabeled and labeled data combined in a single cost function. These Ladder Networks achieve state of the art performance on MNIST (a benchmark dataset about handwritten digits) and CIFAR, a dataset containing 80 million small pictures and their respective class. To further test their abilities, EEG data of the BCI-Wadsworth 2004 dataset was used to build a P300 classifier as a core component of a brain-computer interface. The performance was evaluated with all 64 EEG channels as well as using only 14 EEG channels corresponding to the position of electrodes used by the Emotiv Brainwear-EEG. This thesis shows that Ladder Networks have the potential to increase the performance of a classifier even if labeled data is sparse. Computational efforts aside this thesis concludes that it is possible to achieve state of the art results with Ladder Networks while utilizing only 30% labeled data.

In artificial visuo-haptic object recognition systems, vision and haptics are usually modeled as separate processes. This, however, is far from what actually happens in the human brain, the object recognition capabilities of which we would like such systems to have: A lot of multimodal and crossmodal interactions take place between the two sensory modalities there. Generally, three main principles can be identified as underlying the processing of the visual and haptic object-related stimuli: 1. Hierarchical processing 2. the divergence of the processing onto substreams for object shape and material perception, and 3. the experience-driven self-organization of the integratory neural circuits. The question arises whether an object recognition system can benefit in terms of performance from following a more brain-inspired strategy to combining the visual and haptic input, which we set out to answer in this thesis. We compare the brain-inspired integration strategy with conventionally used integration strategies on data collected with a robot that was enhanced with inexpensive contact microphones as tactile sensors. The results of our experiments involving eleven every-day objects show that 1. the contact microphones constitute a good alternative to capturing tactile information and that 2. the brain-inspired strategy is more robust to unreliable inputs.

Conveying emotional expressions in speech is a key element of human communication. However, until today there is no clear consensus among researchers which prosodic and acoustic features are most relevant for emotional expressions in speech. Many state-of-the-art emotion recognition systems only select the hand-crafted features that achieve a higher classification accuracy in a specific scenario. In contrast to these methods, the aim in this thesis is to contribute to a deeper understanding of the acoustic and prosodic features that are relevant for the perception of emotional states. An artificial deep neural network is proposed that allows to learn the auditory features directly from the unprocessed speech signal. For the analysis of the implicitly learned representations three different visualization methods are utilized: filter visualizations, data-driven visualizations, and network-driven visualizations. On the IEMOCAP dataset, the implemented network has shown to be competitive with state-of-the-art approaches. The results indicate that the network has learned novel features that are complementary to hand-crafted features. The analysis gives detailed insights on the learned representations. For example, the visualizations indicate that higher frequencies, a faster speaking rate, and longer utterances are used for more excited emotional states. Also, clear and constant pauses in an utterance indicate a more negative emotional state. The proposed approach is a general method to enable a deeper analysis and understanding of the most relevant acoustic and prosodic features to perceive emotional expressions in speech.

Almost all available data is unlabeled, so that most current neural networks are not able to take advantage of this data. Deep neural networks achieve state-of-the-art results in many areas by means of better hardware utilization and improvement of the models. However, the main driving factor is the amount of available labeled data. Annotating data always involves human labor and is therefore expensive. The natural conclusion to this is to use semi-supervised models that gain information from labeled and unlabeled data alike. Recently, Ladder Networks have shown promising results in semi-supervised learning of convolutional and feed-forward networks. They reach state-of-the-art results with only a small fraction of the available labels. This thesis introduces Recurrent Ladder Networks that add the benefits of semi- supervised training to domains such as speech and language processing. Speech recognition experiments on the TIMIT corpus show that this model is able to increase the relative accuracy by 10% just by adding unsupervised data. Several variations, including a minimal one that requires little implementation effort but still achieves very similar results, are discussed and evaluated for a comprehensive overview of Recurrent Ladder Networks.

The metric representation of space during navigation is attributed to a number of specialized cells located in the hippocampus and entorhinal corti, which represent and integrate measured distance, direction of motion, speed or any other extra inputs from different parts of the mam- malian brain. Neuroscientific evidence has led to the identification of place cells- these cells fire when an animal is in a specific location; head- direction cells - these fire when an animal faces a chosen direction; and grid cells - these form triangular grid-like patterns that tile the entire environment as an animal moves. Biologically, grid cells are organised into modules where the receptive fields of the cells in one module have the same spacing and orientation (between 30 - 60 degrees) but the scale differs in others forming multiple spatially scaled modules that precisely encode position over a large space. Although the mechanisms through which these multiple spatially scaled modules emerge are still of research interest, existing neural models attribute this modular behaviour to odometry such that the change of the triangular tessellating grid cell firing is influenced by the animal's velocity and direction inputs. In our auto-encoder model, we prescribe to evidence suggesting the existence of auto-associative networks within the entorhinal cortex which cohesively support the emerging activity patterns that spatially represent space during navigation. In this contribution, we present an un-supervised learning auto-encoder model that employs recurrent and temporal sequencing mechanisms of an Elman neural network to enable the emergence of multiple spatially scaled modules of grid cells.

Many modern learning algorithms, such as artificial neural networks, require the practitioner to manually set the values of many hyperparameters before the learning process can begin. However, with modern deep neural networks the evaluation of a given hyperparameter setting takes a lot of time and the search space for the optimization procedure is usually very high dimensional. We propose the usage of a genetic algorithm (GA) to optimize the hyperparameters specifically for convolutional neural networks (CNNs). Additionally, we suggest to use a lower dimensional representation of the original data to quickly identify promising areas in the hyperparameter space. We compare the results of the GA and the hyperparameter optimization on lower dimensional inputs to the Tree of Parzen Estimators (TPE) algorithm, an optimization algorithm based on Bayesian probabilities. Our experiments show that the GA using lower dimensional data representations for earlier generations needs less time to find hyperparameters that perform similar or better than those found by the TPE algorithm. Preliminary results also indicate that under some conditions the GA may be better suited to handle external constraints such as memory limitations. Our results suggest that using genetic algorithms and low dimensional data representations for the earlier generations during the optimization process are a promising way to find appropriate hyperparameters for CNNs and a given problem.

Humans have a remarkable capability of acquiring language and in particular more than one language. More interestingly, they learn it within the same neural computing substrate. But, how the structure of a sentence is mapped to its meaning within the brain is still an open issue. To understand this process of language acquisition, several neural language models have been proposed. Most of the existing models are either using syntactic parse trees to create handcrafted features or using localist vector representation of words to process the sentences. Hinaut et al. proposed the θRARes (Thematic Role Assignment Reservoir) neural language acquisition model, for the Thematic Role Assignment (TRA) task. The model learns the thematic roles, i.e., "Who did what to whom", purely from the grammatical structure of the sentences. The semantic information was removed from the sentences by replacing semantic words with a unique token (`SW'). The model process the transformed sentences considering the words as discrete atomic symbols i.e. using localist word representation. The localist representation of the words do not carry any semantic and syntactic information about the words. This thesis proposes an end-to-end neural language model, Word2Vec-θRARes, in an extension of θRARes model. In addition to utilizing syntactic information and modeling temporal aspect of the input sentences, the proposed model also takes into account the semantics of the words. The model contains a Word2Vec unit to generate the distributed vector representation of words which captures the semantic and syntactic relationships. The model also contains a Recurrent Neural Network namely, Echo state Network (ESN), with fixed random connections, which is capable of modeling the sequential data. Thus, the model receives a raw sentence as an input which is processed word by word across time, and the meaning of the sentence is generated as an output. The results of the experiments conducted using the Word2Vec-θRARes model shows that the exposure of semantic and syntactic information to the ESN increases the performance on the TRA task. Apart from this, the Word2Vec-θRARes model also re-analyse the meaning of the sentence on arrival of words. Also, in some cases, the model correctly predicts the meaning of the sentences even before the sentences are finished.

Elderly people often need some form of assistance during their daily life. A small affordable assistant robot could help by reminding users to take their medicine or by looking for lost items etc. Other uses for such a robot could be in education or HRI research. This work aims at creating a homogenous robotic platform from heterogeneous hardware and software components. These components include a Raspberry Pi 2, an Arduino board, an Android smartphone, a RGB-D camera and the DOCKS speech post-processing. The ROS framework is used to take care of most communication and to enable the potential use of existing ROS modules. Experiments were conducted to proof that the integration of components is working and that the robot possesses basic navigation capabilities. The experiments show that the aim was mostly achieved but the speech recognition and motor control will need some improvements. The resulting robot is a step in the right direction but more work needs to be done for a practical use as assistance robot or for use in education.

Neural coding concerns the neural representation of information. A neural network, whether it is artifical or biological, can encode information using different coding schemes such as rate coding or temporal coding. Applying information theory allows researchers to measure how much information different coding schemes convey. If one coding scheme conveys more information than another then researchers can conclude that this coding scheme is used by the neural network. However researchers have to be careful when applying information theory because estimators are biased. This presentation gives an introduction with examples on how to perform an analysis comparing different neural coding schemes using information theory.

Chromatography is a method from analytic chemistry to purify a mixture of chemical components. Mixtures are processed with a machine and their pure components are extracted. During the analysis, electric signals are captured and recorded. Analyzing data from chemical experiments using Machine Learning algorithms is becoming more and more popular for exploration and intelligent applications due the availability of software tools that can handle large datasets. A method of interest for many applications is to search through a collection of samples and find the most similar one given a new, possibly unknown sample. Chemical software is often based on a traditional way of visualizing the data which makes it difficult to compare two different datasets for similarity. Noise and sparsity make analysis more difficult. To overcome this issues, we explore a novel method for preprocessing and viewing chemical datasets based on a two-step dimensionality reduction method and an image-based representation which combines the classic approaches. Principal Component Analysis (PCA) is the standard method for preprocessing chemical datasets, Linear Discriminant Analysis (LDA) is evaluated as an alternative which performs superior on the underlying dataset. Finally, K-Nearest Neighbors is evaluated for finding the most similar sample using different preprocessed datasets.

The understanding of affective states plays a key role in social relationships between humans. However, it has been shown to be a very challenging task to model in artificial systems. Accordingly, for robots to communicate in a more natural and proactive way, they must respond to human emotional states. In this thesis, a recurrent self-organizing neural architecture that can effectively recognize affective states from human full-body motion patterns was implemented. A dataset called Bodily Expressions of Emotion (BEE) was collected to evaluate the system. This dataset is composed of full-body movements of 6 different emotions: Anger, Fear, Happiness, Neutral, Sadness, and Surprise, that was recorded using a depth sensor. Fifteen participants who did not take part in the data collection were asked to label the data in order to have a human comparison measure. The BEE dataset was used for training and validating the system. Experimental results showed that the system could learn emotional representations from body motion and that the performance of the system was very close to humans performance when recognizing emotions using the same dataset.

Touching is a major part of the human sensory system which is used for expanding the human recognizability of the surrounding environment. Touching is used for recognizing different surfaces and it has a contribution in the proprioception which is represented in the joint position sense. The recognition of surfaces in robots is an important addition to the robot's sensibility of the surroundings. Biologically inspired, a pressure force sensor has been integrated into a robotic arm to classify surfaces using feedback from the sensor only. And this is biologically inspired from the human mechanoreceptors. The sensor determines the surface deformation caused by the contact force with objects. Using the real-time feedback from the sensor, the robotic arm controller and experimental sampling methods were developed. An artificial neural network was used afterward to classify the surfaces and to evaluate the efficiency of experiments in terms of stability, reproducibility and results accuracy.

In this talk we present a multi-modal human robot interaction architecture which is able to combine information coming from different sensory inputs, and can generate feedback for the user which helps to teach him/her implicitly how to interact with the robot. The system combines vision, speech and language with inference and feedback. The system environment consists of a Nao robot which has to learn objects situated on a table only by understanding absolute and relative object locations uttered by the user and afterwards points on a desired object to show what it has learned. The results of a user study and performance test show the usefulness of the feedback produced by the system and also justify the usage of the system in real-world applications, as its classification accuracy of multimodal input is around 80.8%. In the experiments, the system was able to detect inconsistent input coming from different sensory modules in all cases and could generate useful feedback for the user from this information.

The recognition of actions that involve the use of objects has remained a challenging task. We present a hierarchical self- organizing neural architecture for learning to recognize transitive actions from RGB-D videos. We process separately body poses extracted from depth map sequences and object features from RGB images. These cues are subsequently integrated to learn action-object mappings in a self- organized manner in order to overcome the visual ambiguities introduced by the processing of body postures alone. Experimental results on a dataset of daily actions show that the integration of action-object pairs significantly increases classification performance.

A neural network based controller for a humanoid robot arm is designed. The basis of the controller is an inverse kinematics solution, it defines how the robot must move its joints in order to achieve a specified motion of the end effector. This mapping is learned by a multilayer perceptron. Two learning strategies are implemented, optimized, and compared: goal and motor babbling. Motor babbling relies on random motor commands while the more recent goal babbling approach uses goal directed motions. Each of these methods is employed in both a three-dimensional and a six-dimensional task space which includes the orientation of the end effector. Additionally, a solution for relative inverse kinematics is implemented and tested. These are finally tested and compared using the NICO robot.

The latest change of rules for the RoboCup Soccer in the humanoid league allows the balls' surface to consist of up to 50% of any color or pattern, whereas at least 50% have to be white. This makes ball localization an even more important and challenging task, because multi-color balls have changing color histograms and patterns, depending on the balls' current orientation and movement. To handle this newly introduced difficulty this thesis proposes a neural architecture, a convolutional neural network (CNN), to localize the ball in various scenes. CNNs are capable of learning invariances in images and are used in several image recognition tasks. In this thesis, CNNs are designed to locate a ball by training two output layers, one for the x- and one for the y-coordinates. The CNNs get trained with normal distributions fitted around the ball center. This makes it possible for the network to not just locate the ball position but also to provide an estimation of the noise. The whole image is processed by the architecture in full size, no sliding-window approach is used. Afterwards, the CNNs output gets filtered by a recurrent neural network (RNN), which additionally tries to predict future positions of the ball.

Our understandings of the behavioural and mechanistic characteristics for natural language are still in its infancy and we need to bridge the gap between the insights from linguistics, neuroscience, and behavioural psychology. To contribute an understanding of the appropriate characteristics in a brain-inspired neural architecture that favour language acquisition, recurrent neural models have been developed for embodied and multi-modal language processing, embedded in a developmental robotics framework. In this disputation the main contributions from the study of these models are reported.

Robots in domestic environments are receiving more attention, especially in scenarios where they should interact with parent-like trainers for dynamically acquiring and refining knowledge. A prominent paradigm for dynamically learning new tasks has been reinforcement learning. However, due to excessive time needed for the learning process, a promising extension has been made by incorporating an external parent-like trainer into the learning cycle in order to scaffold and speed up the apprenticeship using advice about what actions should be performed for achieving a goal. In interactive reinforcement learning, different uni-modal control interfaces have been proposed that are often quite limited and do not take into account multiple sensor modalities. In this paper, we propose the integration of audiovisual patterns to provide advice to the agent using multi-modal information. In our approach, advice can be given using either speech, gestures, or a combination of both. We introduce a neural network-based approach to integrate multi-modal information from uni-modal modules based on their confidence. Results show that multi-modal integration leads to a better performance of interactive reinforcement learning with the robot being able to learn faster with greater rewards compared to uni-modal scenarios.

In the future, robots are expected to work as a companion with humans in various areas ranging from service robots to humanoid robots. Dynamic and unpredictable human/domestic environments force developers to improve safety for human-robot cooperation. One natural approach for humans is to warn about threats using natural spoken language. Then, robots should be able to modulate safer actions by a syntactic/semantic understanding of those warnings. In pour research deep neural networks will be used as the main learning approach of the natural language processing part. However, besides warning messages, other modalities seem necessary to gain a better understanding of threats such as prosody and vision. Generating safer actions depending on context can be performed by reinforcement learning or simply by choosing from an available action set. Moreover, possible tasks and scenarios as well as datasets and platforms will be discussed in this talk.

When robots interact with humans by verbal dialogs, then the question arises whether the robot could predict dangerous situations or actions through language processing. As the robot demand in the area of domestic service increases, the safety in human-robot interaction becomes crucial. It could be beneficial, that the robot could learn safe and unsafe actions, to prevent the dangers that can occur in an uncertain environment. This talk will focus on exploring or generating the corpora for learning safe and unsafe concepts, and I will propose the utilization of annotated or unannotated corpora in a natural language interface. I already used a rule-based verbal dialogue system, Yet Another Dialogue Tool Kit (YADTK), to build a verbal interface. The PhD research under the SECURE project will be focused on inferring and learning of safety concepts primarilybylanguage alone, and will be tested on the NAO humanoid robot platform. We will discuss different possible corpora and datasets, with possible language models and planned experiments.

Class imbalance is a problem in machine learning often found in real world datasets where one of the class dominates entire dataset. Most machine learning algorithms fail in classification of such dataset if class to class separability is poor. This thesis aims at finding methods to improve performance on poorly separated imbalanced datasets using Cost Sensitive Neural Networks. The focus dataset which have these characteristics is from XING AG and the problem is to predict a user's job change in near future. Results show that despite being imbalanced and poorly separated, performance metrics such as G-Mean as high as 92.8\% could be reached by using cost sensitive Convolutional Neural Networks to find patterns in time series. This thesis also generalizes the methods used and results show that it can be applied to any imbalanced dataset.

Pattern recognition is an active field of research in the artificial neural networks community. In this thesis an artificial neural network, the so called Echo State Network (ESN) will be utilized to recognize ten di?erent hand gestures recorded by the inertial measurement unit of a regular smartphone. There are many different approaches for gesture recognition, mostly consisting of a segmentation and a classification phase. We will give a short overview over the different approaches and then present our own framework, from data collection to the finished classifer. We show how to set up the environment of the ESN to allow for good classi?cation results. As we use real world data the samples are of different length and power. Additionally collected sensor values contain noise and bias. We will show that the nature of the ESN not only allows to deal with these problems, but also others the possibility to perform segmentation and classification at the same time. We achieved classification accuracies as good as comparable approaches that use complicated segmentation algorithms by utilizing only a single artificial neural network.

The field of neurocognitive robotics takes the processing mechanisms of the brain as inspiration and guidance: computer implementations of robot perception and action should be based on brain-like neural architectures and biologically plausible learning mechanisms. Unsupervised learning and reinforcement learning have led to good results on the emergence of internal sensory representations and intelligent reward-seeking behaviours, respectively. However, other aspects of animal behaviour are generally not considered, even though it has been argued that only a more comprehensive study of animal behaviour can lead to a deeper understanding of intelligent behaviour. This thesis does not attempt to provide a comprehensive model of animal behaviour, but rather tries to draw attention to the need for it by presenting the potential of neglected aspects of animal behaviour such as self-preservative behaviour.

Self-preservative behaviours are believed to impose the ground rules for more complex and motivated behaviour. Although many of these innate responses are hard-coded in the brain, they are not sufficient for the organisms' survival. They have to adapt, by learning, to new and unexpected situations within their lifetime and thereby be able to interact effectively with their environment. A key component on the lifetime adaptation is the formation of associations/memories between environmental predictors and relevant events, which mainly rely on punishment and reward learning.
We believe that a deeper understanding of innate and learned defensive mechanisms could also be helpful in developing future robot generations, making them more adaptable and robust. Therefore, in this thesis, we study and develop three neurocomputational self-preservative mechanisms in the context of humanoid service robots to demonstrate the potential and feasibility of including bio-inspired adaptive self-preservative mechanisms as part of real-world robotic systems. Our aim is to present possible ways in which robots can be endowed with such adaptive self-preservative mechanisms at different neurocognitive levels, going from abstract biological models to neurocomputational models.

The first experiment addresses the problem of search for an appetitive stimulus. Here a reinforcement learning (SARSA) algorithm was optimized to learn in a real-world scenario and manoeuvre a humanoid robot towards a charging station. The second experiment focuses on the role of punishment and nociceptive sensory input in motor learning. Both types of feedback play an important role in driving attention, and modulating decision making and action. However, they have not been thoroughly studied in computational models. Here, we compared the effect of both types of feedback on an Actor-Critic learning algorithm (CACLA). Finally, in our last experiment, we studied the role of noxious stimuli in the formation of anticipatory behaviour. This experiment is based on Pavlovian and instrumental conditioning and how environmental cues can be used to anticipate negative outcomes. A hybrid approach using an echo state network (ESN) and a dopamine modulated Pavlovian conditioning model was used to anticipate nociceptive sensory input based on auditory cues.
In all three experiments we showed how often neglected, self-preservative mechanisms could solve meaningful artificial intelligence problems while providing the basis for new neuro-inspired computational processes. In particular, we showed how bio-inspired sensorimotor signals associated with nociception and pain can be exploited for learning beyond triggering reactive behaviours. We also developed novel extensions to the learning algorithms used.

What if the robot could predict dangerous situations in the home environment? It could be very beneficial in various aspects: the robot could be aware of its behaviour and prevent potential harm to a person or it could signal if there is a threat to a person, which might save lives.
In this talk I will present a research project called "Identifying Dangerous Situations from Acoustic Data using Prosody Features" as a part of the SECURE project. The goal is to teach a robot to identify threatening situations, for instance, when a person screams for help.
My current PhD research will be focused on acoustic and speech processing by means of deep neural networks with exploiting unlabelled data and testing them on the iCub robotic platform. In this talk, the background of the project will be presented and the possible datasets and experiments will be discussed.

Games have always been traditionally a research area for artifcial intelligence. Through them, a platform is given to directly test computational capabilities against human cognitive abilities and knowledge in a certain domain. Poker is currently the most played card game and one of the few games left where professional human players play at a higher level then computer players. This is mainly because of the imperfect information character of this game. To handle this issue, several basic directions emerged in research on this topic over the last twenty years. One of them is the exploitive counter strategy. The aim of this strategy is to fill the missing information gap by incorporating a model of the opponent and moreover, to exploit his suboptimal game play with aid of this additional knowledge. Although Poker is undeniably a game where the decision for a certain action is heavily influenced by previous game events, all found publications so far used machine learning approaches, being suboptimal for temporal tasks, when modelling the opponent. In this thesis we introduce Echo State Networks for modelling a players' strategy in Texas Hold'em Poker. Echo State Networks are Recurrent Neural Networks following the Reservoir Computing paradigm. As such, they are not only able to detect patterns in temporal dependencies, but are also provided with effective and fast training procedures.

Automated speech recognition (ASR) deals with the processing of spoken audio-data into processable data. The hybrid ASR-system DOCKS uses Googles "Search by Voice" as cloud-based ASR-system to receive a text-hypotheses for an audio-input and Sphinx-4 as local ASR-system to improve the result with domain-specific language models and a restricted vocabulary. Currently a bi-gram language model is the highest-order n-gram model for the DOCKS system. In this thesis a tri-gram language model for the hybrid ASR-system DOCKS is presented and the performance of this approach was measured by evaluating it on different data-sets. The results of the experiments show, that the performance of the hybrid ASR-system is highly depended on how well the domain of the inputs is covered by the language-model. For a highly restricted domain with n-gram models that sufficiently cover it, the tri-gram language model post-processing approach presented in this thesis achieves the best performance compared to other local, cloud-based and hybrid ASR-systems.

We present a neural probabilistic robot localization and navigation model that provides a mere concept while enabling far-reaching functionality. Through emulating several basic functionalities of the brain the system is able to achieve complex tasks such as robust target tracking, environment learning through observation, and flexible robot navigation in a home-like environment. The concept of our work is implemented and evaluated using a robot platform in a home-like environment, and the results show that our neural system helps a robot to realize different functions successfully.

Gradient-based learning in Recurrent Neural Networks (RNNs) is difficult due to the vanishing gradient problem. This problem arises when RNNs have to be unfolded in time, making them too deep for efficient backpropagation. However, this temporal depth is required to effectively capture long-term dependencies in sequences. Multiple solutions have recently been proposed that show promising improvements and a better understanding of the issues at hand. Along these proposals are RNN models that allow the emergence of multiple timescales during training. To contribute to the understanding of the fundamental conditions that allow RNNs to self-organize to rich temporal dynamics, a number of these models are explored on sequence generation and prediction tasks. We compare the Simple Recurrent Network (SRN) with the recently proposed Clockwork RNN (CWRNN), Structurally Constrained Recurrent Network (SCRN), and the Recurrent Plausibility Network (RPN). The timescaling mechanisms of the SCRN and the RPN are adapted to the CWRNN. The widely used Long Short-Term Memory (LSTM), which is specifically designed to deal with vanishing gradients, is included as well.

In his Bachelor thesis Stefan Bruhns has created a new network type called 'modulated spiking neural network'. This thesis expands his work by analysing the performance of the new network on different tasks and comparing them to different other network types.

The integration of multisensory information plays a crucial role in autonomous robotics. In this talk, I will vividly argue that robust multimodal representations can naturally develop in a self-organized manner from co-occurring multisensory inputs. In particular, I propose a hierarchical learning architecture with growing self-organizing neural networks for learning human actions from audio-visual inputs. The hierarchical processing of visual inputs allows to obtain progressively specialized neurons encoding latent spatiotemporal dynamics of the input, consistent with neurophysiological evidence for increasingly large temporal receptive windows in the human cortex. Associative links to bind unimodal representations are incrementally learned by a semi-supervised algorithm with bidirectional recursive connectivity. Multimodal representations of actions are obtained using the co-activation of action features from video sequences and labels from automatic speech recognition.

Along with the ongoing progress in robotics, human robot interaction becomes increasingly important. The EchoRob project aims at natural language acquisition in robots, based on an artificial neural network. It therefore needs a huge amount of language data from the context of human robot interaction in form of texts, speech and annotations. To obtain such data is challenging and time consuming. That is why in this thesis I introduce a design approach for a crowdsourcing website that enables laypersons to help generating that data on a website. I not only suggest the design for a specific set of tasks that should lead to high quality data but also a gamification system that provides important feedback to the users and keeps them motivated during the long run. My design suggestions are based on findings from motivational psychology, research on other citizen science projects and a PACT analysis of the planned website.

The thesis deals with the Stable Marriage Problem (SMP) by reference to the Student Topic Assignment (STA) at Universities. In the SMP one tries to achieve stable mappings between two groups which have priorities for each other. A valid mapping is accomplished, if a bijection 1:1 between these groups is reached. While there is already an algorithm for solving the SMP by Gale and Shapley there is no way to get the best, but still invalid solution for the problem, if the SMP is not solvable. The Particle Swarm Optimization (PSO), originally proposed by Kennedy, Eberhart and Shi, is a swarm-intelligence-based algorithm. It is utilized to solve optimization problems like the STA. The peculiarity in the student to topic assignment is that it seems to be an instance of the Stable Marriage Problem, but rather is in the class of Assignment Problems (AP). Discussing this controversy and solving the STA, this thesis will show four PSO variations developed for this purpose.

We used visualisations to analyse CNNs trained on facial expression recognition by visualising neurons or feature maps. Through the visualisations we are able to understand the decision-making process and modify the databases and the architecture as needed. As a consequence, we learned that more sophisticated arrangements of classes allows filter maps to learn more differentiated information about the domain. Training CNNs on top of these data representations will improve classification results for all other classification tasks in the domain. We conduct multiple experiments directed at different parts of the architecture, through which we illuminate the mechanisms of CNNs and in this way facilitate the understanding of them. Furthermore, analysis through visualisations allows us to detect deficiencies of the architecture apart from the commonly known phenomenon of ``overfitting'', which might be the first thought that comes to mind, especially in case of generalisation to other databases. These include the suitability of the distribution of images throughout the classes, which can not be easily detected.

Hierarchical reinforcement learning solves complex reinforcement learning tasks by dividing them into subtasks and solving them individually in order to solve the whole task. This thesis presents some approaches that try to solve a complex reinforcement learning problem hierarchically using biologically inspired neural networks. The currently available hierarchical reinforcement algorithms are able to solve complex tasks hierarchically once the hierarchy has been defined by the programmer but generally are unable to find any hierarchy in the tasks themselves.
The goal of this thesis, in addition to solving a complex learning task, is to discover hierarchy in the task. The thesis then presents the architectures for each approach and how they performed. The results of these approaches show that overall they were able to find some structure, although not hierarchical, and solve the defined task hierarchically.

Motivated by the efficiency of Convolutional Neural Networks in implicit feature extraction and the ability of Long Short-Term Memory Recurrent Neural Networks in learning long-range time dependencies, this thesis proposes a Convolutional Long Short Term Memory Recurrent Neural Network (CNNLSTM) for the task of gesture recognition. The proposed system requires as input a stream of differential images, each of them representing the extracted body motion. Each of the differential images is processed by the proposed CNNLSTM architecture, in order to be assigned a gesture label. CNNLSTM consists of two consecutive Convolutional layers, the flattened output of which goes through a Long Short-Term Memory Recurrent layer. The output of the recurrent layer is fed to a softmax output layer. CNNLSTM is a deep recurrent neural network model and therefore can be trained with Backpropagation Through Time (BPTT). The performance of the system is tested on three different datasets and proves significantly higher than the performance of a baseline feedforward Convolutional Neural Network. CNNLSTM also exhibits common learning behaviour in all the tested datasets independently of the difficulty of each dataset.

Gestures are one of the main communication channels in Human-Robot Interaction (HRI). Because every individual performs gestures differently, personal robots need to be able to adapt to various persons and environments, while efficiently managing their computational resources. In this thesis, we present an approach to realize gesture recognition with Echo State Networks in an online learning scenario. Echo State Networks are a special type of Recurrent Neural Networks, but have the advantage of a more power effective training. We introduce an efficient preprocessing method to prepare image sequences for the network, which is then trained by using a mini-batch gradient descent algorithm to change the network's outputs over time. The recognized gestures are then used to control an iCub robot head for HRI. We show that Echo State Networks can achieve a good performance on the given test data and reach a decent level of adaption in a live scenario, where new unseen data is presented to the robot by a new subject.

The superior colliculus is a mid-brain region which integrates sensory input to localize mul-tisensory stimuli. This thesis aims to close the gap between models of SC physiology and system-level behavior, and development thereof. A new model of the SC based on self-organized statistical learning is proposed and it is shown to replicate a variety of important biological phenomena. It is then applied to a problem in robotics: binaural sound-source localization. We show that our algorithm can learn to per-form state-of-the-art sound-source localization.

The study of emotions has attracted considerable attention in several areas, from artificial intelligence and psychology to neuroscience. The use of emotions in decision-making processes is just one example of how multi-disciplinary the research in emotions are.
In my work I assume that communication in a Human-Robot interaction scenario, for instance with gestures, could benefit from additional human emotion recognition. In this presentation, I will show a deep neural network model which is able to recognize spontaneous emotional expressions and to classify them into 3 categories: positive, negative or neutral. The model was evaluated in two experiments, one using benchmark datasets, and the other using an HRI scenario with a humanoid robotic head, which itself gives emotional feedback.

You already have tweaked some (hyper-)parameters to improve the performance of your algorithm? Then this tutorial is for you: "|hyperopt| is a Python library for optimizing over awkward search spaces with real-valued, discrete, and conditional dimensions."
It enables you to make your hyper-parameter search more automatic and reproducible than if you just do by "trial and error". The library comes with a lot of useful functionalities that you do not have to implement yourself (e.g. Random or Tree of Parzen Estimators (TPE) search algorithms). In particular, you can define your search space on a very fine-grain level of details: continuous, discrete, or logarithmic, and you can have conditional parameters (some may exist or not depending on other parameters), which means you can search over different type of search spaces at the same time. This tutorial will take a look on the basics and trying to get hints on the advanced parts. Follow the link hyperopt for further information. For any questions left, join us and make the tutorial a dynamic exchange for all of us. If you want to know other existing sub-librairies (ConvNets, scikit-learn, etc), click here

It is known from studies in the monkey ventral pathway that two-dimensional pattern representations are processed in areas such as V2, V4 and IT cortex. However, the role played by the brain to process three-dimensional shape is unclear. Some recent studies suggest that IT neurons encode also three-dimensional spatial configurations, useful for understanding objects, for example. This hypothesis is consistent with classical theories about shape representation based on the concept of Geons. In this work we will presented a neural network designed to extract three-dimensional features from point cloud data for object classification. At the same time, the importance of multiple view points in the recognition, largely studied in psychology and confirmed by many computational studies, is taken into account. Experiments using the Washington Dataset composed of 51 categories and 300 instances of objects shows an improvement using three-dimensional features and motivates further investigation.

As robots have to do more and more human tasks, they need more human-like moving capabilities. In this thesis, a biologically-inspired joint operating with tendons is built. The controlling for such joints is examined on this example. It is tried to control the joint with learning algorithms like goal babbling and neural networks. On the one hand, tendons have some advantages over motors directly in the joint as they are space-saving when applied to fingers or can protect the motor from the forces if the robot falls down. On the other hand, controlling a joint operated by tendons is much harder than a simple servo controlled by a PID-Controller.

The use of emotional states for Human-Robot-Interaction (HRI) scenarios has attracted considerable attention in recent years. One of the most challenging task is to recognize the spontaneous expression of emotions. Every person has a different way to express emotions, and this is aggravated by the complexity of interaction with different subjects, multi-cue information and different environments. How to deal with the particularities of each person, each expression and each cue? In this work, I will show how our convolutional neural network based architecture learns features that can represent these particularities, and which are the advantages of using them to recognize spontaneous emotions.

The main motivation of this work is to gain insight into improving the reliability of color information by following the coding strategies of the lower stages of the human visual system. In comparison to human perception, the color information of a camera is unreliable as it is greatly in uenced by the luminosity and color of the illumination of a scene. Established unsupervised computational models of self-organizing RGC and V1 receptive fields are extended to process RGB images and are trained upon images of natural scenes. The results show that localized color opponent RF are emerging in chromatically distinct channels, each entirely covering the visual space. The opponent texture of the resulting RF is filtering more biologically plausible chromatic contrasts, which are regarded as the fundamental building blocks of the human visual system's property of color constancy. Because of the unmatched performance the visual system produces, filters capturing more biological plausible chromatic contrasts appear to be of value and are suggesting that computational models utilizing such preprocessing possibly show an improved performance.

The correct execution of well-defined movements in sport disciplines may increase the body's mechanical efficiency and reduce the risk of injury. While there exists an extensive number of learning-based approaches for the recognition of human actions, the task of computing and providing feedback for correcting inaccurate movements has received significantly less attention in the literature. We present a learning system for automatically providing feedback on a set of learned movements captured with a depth sensor. The proposed system provides visual assistance to the person performing an exercise by displaying real-time feedback to correct possible inaccurate postures and motion. The learning architecture uses recursive neural network self-organization extended for predicting the correct continuation of the training movements. We introduce three mechanisms for computing feedback on the correctness of overall movement and individual body joints. For evaluation purposes, we collected a data set with 17 athletes performing 3 powerlifting exercises. Our results show promising system performance for the detection of mistakes in movements on this data set.

Recently robots are being used more frequently as assistants in domestic scenarios. In this context we train an apprentice robot to perform a cleaning task using interactive reinforcement learning since it has been shown to be an efficient learning approach benefiting from human expertise for performing domestic tasks. The robotic agent obtains interactive feedback via a speech recognition system which is tested to work with five different microphones concerning their polar patterns and distance to the teacher to recognize sentences in different instruction classes. Moreover, the reinforcement learning approach uses situated affordances to allow the robot to complete the cleaning task in every episode anticipating when chosen actions are possible to be performed. Situated affordances and interaction allow to improve the convergence speed of reinforcement learning, and the results also show that the system is robust against wrong instructions that result from errors of the speech recognition system.

This thesis focuses on discovering the effect of the ship's surrounding environmental conditions on the speed and fuel consumption, which contribute to the final cost of a planned voyage. This thesis is supported by Skysails GmbH, the world famous developer of the SkySails kite propulsion system. Skysails produces and distributes a marine performance management tool that operates in many cargo ships. Therefore, a huge amount of ambient data have been acquired from different vessels that operate on harsh offshore environment. In this thesis, we are investigating whether the regression methods (neural networks, support vector regression) are able to derive fuel consumption parameters for a future voyage of travelling ship or more generally learning a true function from ambient data.

Reservoir Computing (RC) is an attractive training mechanism for recurrent neural networks due to its ease and has therefore gained popularity in the neural network community. In this talk, we will present our approaches to RC addressing the Echo State Networks applied to gesture recognition and semantic parsing. We will present our experimental set ups and show recent results from our research.

In this talk I will briefly present my background and motivation, preliminary work done recently and the outline of my research project. The ongoing development of robotics and the growth in number of elderly individuals have led to an increasing interest on applying assistive robotics in order to maintain or even improve the quality of care. At the hospital, in elderly care centres and domestic environments a patient can be assisted by making sure he/she performs essential daily activities such as eating, drinking and taking medications. Therefore, thorough understanding of the human-object interaction activities is the most important building block of an effective assistive framework. Our focus is on domestic scenarios in which human activities involve humans and objects of common use. The recognition of human-object interaction activities requires the analysis of the objects present in the scene, the analysis of human actions, and finally the analysis of interplay between objects and human actions. Therefore, our goal is the implementation and evaluation of a neural-network-based framework, able to learn, represent and integrate two different perception tasks: object recognition and human action recognition.

In reinforcement learning, the options framework allows us to implement intelligent agents capable to learn temporal abstractions. Such framework is based on the theory of semi-Markov decision processes. In this talk, I will give an introduction to the options framework and I will discuss some research lines that can be followed to investigate the design of intelligent agents capable to perform both: multi-sensory integration and temporal abstractions learning by means of options.

Short Bio
Victor Uc-Cetina received a PhD degree (Dr. rer. nat.) in Computer Science from Humboldt-Universität zu Berlin, Germany, in 2009, a M.Sc. in Intelligent Systems from the Monterrey Institute of Technology and Higher Education, Mexico, in 2002, and a B.S. in Computer Systems Engineering from the Merida Institute of Technology, Mexico, in 1999. He is currently a professor in the Faculty of Mathematics at the Autonomous University of Yucatan, in Mexico. His research interests are mainly focused on artificial intelligence and machine learning methods.

Statistical n-gram language models are widely used for their state of the art performance in a continuous speech recognition system. In a domain based scenario, the sequences vary at large for expressing same context by the speakers. But, holding all possible sequences in training corpora for estimating n-gram probabilities is practically difficult. Capturing long distance dependencies from a sequence is an important feature in language models that can provide non-zero probability for a sparse sequence during recognition. A simpler back-off n-gram model has a problem of estimating the probabilities for sparse data, if the size of n-gram increases. Also deducing knowledge from training patterns can help the language models to generalize on an unknown sequence or word by its linguistic properties like noun, singular or plural, novel position in a sentence. For a weaker generalization, n-gram model needs huge sizes of corpus for training. A simple recurrent neural network based language model approach is proposed here to efficiently overcome the above difficulties for domain based corpora. A neural probabilistic language model is directly integrated with the cloud based speech recognition system for its real time usage rather than the common two-pass n-best list rescoring method. Probabilistic n-gram and neural network language models are evaluated on the TIMIT Core Test and Scripted Robot Command corpus for their overall performance based on word error rate and precision. The experiments show that both the language models perform equally well with no significant difference in word error rate, but recurrent neural network model exhibited generalization and capturing long term dependencies slightly better than n-gram model for the corpora.

A way of neuromodulation of spiking neurons and its impact to artificial neural networks is shown. In this context the difference to networks with spiking neurons without neuromodulation and known GasNet model are taken into account. A Shape-Discrimination-Task is used for comparison. In that task a robot should recognize two different objects based on its shape. Neuromodulation offers more ways to map dynamics within a neural network.

The brain comprises hierarchical modules on various physiological levels. Neural feedback signals modulate the neural activities via inhibitory or excitatory connections within/between these levels. They have predictive and filtering functions on the neuronal population coding of the bottom-up sensory-driven signals in the perception-action system. In this thesis, we propose that the predictive role of the feedback pathways at most levels of action and perception can be modelled by the recurrent connections in different artificial cognitive platforms. This will be examined by three recurrent neural network models. Furthermore, the three models and experiments with them show that the recurrent neural networks are able to model feedback pathways and to exhibit the feedback-related sensorimotor predictive functions.

Growing theoretical and experimental research on action and language processing and on number learning and space representation clearly demonstrates the role of embodiment in cognition and language processing. In psychology and neuroscience this evidence constitutes the basis of embodied cognition, also known as grounded cognition (Pezzulo et al. 2012). In robotics, these studies have important implications for the design of linguistic capabilities in cognitive agents and robots for human-robot communication, and have led to the new interdisciplinary approach of Developmental Robotics (Cangelosi & Schlesinger 2015). During the talk we will present examples of developmental robotics models and experimental results from iCub experiments on the embodiment biases in early word acquisition studies, on word order cues for lexical development and number and space interaction effects. The presentation will also discuss the implications for the "symbol grounding problem" (Cangelosi, 2012) and how embodied robots can help addressing the issue of embodied cognition and the grounding of symbol manipulation use on sensorimotor intelligence.

References
Cangelosi A. (2012). Solutions and open challenges for the symbol grounding problem. International Journal of Signs and Semiotic Systems, 1(1), 49-54 (with commentaries) Cangelosi A, Schlesinger M (2015). Developmental Robotics: From Babies to Robots. Cambridge, MA: MIT Press.
Pezzulo G., Barsalou L.W., Cangelosi A., Fischer M.H., McRae K, Spivey M.J. (2011). The mechanics of embodiment: a dialog on embodiment and computational modelling. Frontiers in Psychology, 2(5), 1-21

Biography
Angelo Cangelosi is Professor of Artificial Intelligence and Cognition and the Director of the Centre for Robotics and Neural Systems at Plymouth University (UK). Cangelosi studied psychology and cognitive science at the Universities of Rome La Sapienza and at the University of Genoa, and has been visiting scholar at the University of California San Diego and the University of Southampton. Cangelosi's main research expertise is on language and cognitive modelling in humanoid robots, on language evolution in multi-agent systems, and the application of bio-inspired techniques to robot control (e.g. swarm of UAVs). He is the coordinator of the Marie Curie ITN "RobotDoC: Robotics for Development of Cognition" (2009-2014) and the UK EPSRC project "BABEL: Bio-inspired Architecture for Brain Embodied Language" (2012-2016), and of the FP7 project "ITALK" completed in 2012. Cangelosi has produced more than 200 scientific publications, and has chaired numerous workshops and conferences including the IEEE ICDL-EpiRob 2011 Conference (Frankfurt, August 2011). In 2012-13 he was Chair of the International IEEE Technical Committee on Autonomous Mental Development. He will be Editor-in-Chief of the IEEE Transactions from Autonomous Development from January 2015, and currently is Editor of the journal Interaction Studies. He is co-author, with Matt Schlesinger, of the forthcoming book "Developmental Robotics: From Babies to Robots" (MIT Press, 2015).

Abstract: Humans learn to adapt to high complex, high dimensional real world problems. The separation of neural circuits into modules, which act on disjunct functions, enables the human to adapt. Dopamine in the basal ganglia encodes differences of expected and experienced reward. External stimuli do not only occur in positive matter, but also appear as negative signals. The amygdala represents strong negative external stimuli as fear. The nuclei complex adapts fast to fear experiences. Over time, discrimination takes place and the fear signals are understood more precisely. There are two types of dopamine releases that reinforce learning. Phasic mid brain dopamine fires on positive external stimuli. Tonic dopmaine is suppressed on negative external stimuli. The phasic activation, from positive stimuli, is unproportionally higher than the suppression of tonic dopamine. As a result a "dopamine-ramp" occurs. The ramp disadvantages learning on negative stimuli in the basal ganglia. The abstraction of dopamine as temporal differential (TD) error is used to teach an agent to navigate towards a state with reward. As well as for reward in navigation, artificial neural networks can be used to map input signal to states with fear. The artificial neural network maps input signals to Q-values, which encode volitional behavior as an answer of the input signals. Contrary external stimuli, reinforced on one neural structure, cause interference. An artificial multi layer perceptron (MLP) shows no behavior adaptation to solutions that are reinforced with contrary signals. An agent, controlled by a single structure MLP, is not capable to learn the mapping of input signals to actions. The positive TD error for reward contradicts the fear reinforcement. As a result of this interference, the agent does not learn the non-linear approximation of the input-ouput mapping. As a remedy to this problem, two separate neural circuits can have distinct weight changes via TD error for pleasant reward and TD error for fear. Each neural circuit of the separated MLP has half the amount of neurons in the hidden layer. The separation into two neural circuits, show disjunct stepwise optimization towards fear and reward learning. The disjoined structure shows adaptation to conjugate external stimuli. In practice, the discretization of neural structures into disjunct functions, is not proportional to the complexity of the minimizing error function. Contrary action selections of the agent extend exploration.

Speech-based human-robot communication is natural and efficient but also very prone to error. Therefore, the purpose of this bachelor thesis was the development of a multimodal Android interface for ROS-based robots. The application connects to a ROS master and awaits closed-ended question requests. Users can either speak to the robot or use the touchscreen to answer these questions. To address the issue of poor status visualisation in existing interfaces the application utilises an avatar, animated graphics and icons. User tests have shown that the provided solution is capable of delivering resilient communication in common scenarios. The results of the speech recognition have been improved by the post-processing software DOCKS. Errors in the speech recognition were handled gracefully by the touchscreen.

Giving interactive feedback, other than well done / badly done alone, can speed up reinforcement learning. However, the amount of feedback needed to improve the learning speed and performance has not been thoroughly investigated. To narrow this gap, we study the effects of one type of interaction: we allow the learner to ask a teacher whether the last performed action was good or not and if not, the learner can undo that action and choose another one; hence the learner avoids bad action sequences. This allows the interactive learner to reduce the overall number of steps necessary to reach its goal and learn faster than a non-interactive learner. Our results show that while interaction does not increase the learning speed in a simple task with 1 degree of freedom, it does speed up learning significantly in more complex tasks with 2 or 3 degrees of freedom.

Nowadays, many humanoid teen sized robot platforms have been developed by different research groups. The idea is either to fulfill the research purposes or to produce a specific task fulfilling machine. This imposes many challenges on the design of algorithms for different actions like walking or reaching some target. There are many sophisticated humanoid research platforms available, but one crucial thing to look at is the developmental cost associated with them. As the name describes, humanoid robots are the robots that resemble humans in their design as well as their way of performance. This thesis aims at contributing to the arm design for humanoid robots and to the learning to approaching a target. We used 3D plastic printing for part manufacturing. An arm with multi degrees of freedom is designed to enable the robot free movement around the body. This thesis also contributes in the field of designing a simulator model of a robot. The research carried out in this thesis takes the early learning in infants as the basis. We used this idea to develop a learning algorithm that eventually enable the robot to reach goals in 3D space with accuracy.

Primates can learn complex sequences that can be represented in the form of categories, and even more complex hierarchical structures such as language. The prefrontal cortex is involved in the representation of the context of such sequential structures, and the striatum permits the association of this context with appropriate behaviours. In this experiment, the Reservoir Computing paradigm is used: a recurrent neural network with fixed random connections models the prefrontal cortex, and the read-out layer (i.e. output layer) models the striatum. The model was applied to language syntactic processing, especially thematic role assignment: given a sentence this corresponds to answer the question "Who did what to whom?". The model processes categories (i.e. abstractions) of sentences which are called "grammatical constructions". The model is able to (1) process correctly the majority of the grammatical constructions that were not learned, demonstrating generalization capabilities, and (2) to make online predictions while processing a grammatical construction. Moreover, we observed that when the model processes less frequent constructions an important shift in output predictions occurs. It is proposed that a significant modification of predictions in a short period of time is responsible for generating Evoked-Related Potentials (ERP) like the P600: this typically occurs when unusual sentences are processed. The use of the model for complex abstract sequence processing shows that the faculty of representation and learning of sequences in the brain may be based on highly recurrent connections. This experiment suggests that artificial recurrent neural networks can provide insight into the underlying mechanisms of human cortico-striatal function in sentence processing. Finally, to show the ability of the model to deal with a real-world application, the model was successfully applied in the framework of human-robot interaction for both sentence comprehension and production. The sentence production model was obtained by "reversing" the sentence comprehension model: it processes meanings as input and produces sequence of words in output. More recently, an incremental version of the model with Hebbian-like learning has also been developed, demonstrating the biological plausibility of the underlying mechanisms. This incremental version could be used in a developmental learning scheme for robots.

Certain common problems in pattern detection may be regarded as timescale re- lated, e.g. detection of patterns of varying lengths, slightly (time-)warped variations of same patterns and short-scale variations of pattern values, such as noise. This thesis investigates those problems in the context of pattern detection with 'Echo State Networks' (ESNs), a reservoir-based type of recurrent neural networks. It aims to look at ESN behaviour and to shed light on ESN parameter in uence on good performance when confronted with those problems. We conduct experiments on synthetic data and optimize different ESN models using evolutionary optimiza- tion in the context of timescale dependent problems. We then analyse behaviour and parameter distributions.

The thesis analyses different supervised text categorization methods to find optimal feature vector sizes for given application scenarios. We use feature selection algorithms to reduce the dimensionality of the input data and supervised classifiers for the actual categorization task. Several metrics are used to evaluate and to compare the different approaches. Based on the results suggestions are given for the applications scenarios. As a basis the vector space model is used. Feature selection is done by Chi-Square, information gain and oneR. Each of them is combined with the classifiers, we used supervised self organizing maps (SOM), decision trees and the support vector machine (SVM). The result suggest that self organizing maps with information gain are more suitable for time-constrained applications, where information gain in combination with support vector machine is better for best performance with no time constraint. In general, SVMs are performing very well and decision trees are very efficient in classification. This thesis also gives a guideline to analyse performance with different feature vector sizes for future work.

In the recent past robots are being used more frequently as assistants in domestic scenarios. In this context we train an apprentice robot to perform a cleaning task using interactive reinforcement learning since it has been shown to be an efficient learning approach benefiting from human expertise on performing domestic tasks. The robotic agent obtains interactive feedback via a speech recognition system which is tested to work with five different microphones concerning their polar patterns and distance to the teacher to recognize sentences in different instruction classes. Moreover, the reinforcement learning approach uses situated affordances to allow the robot to complete the cleaning task in every episode anticipating when chosen actions are possible to be performed. Situated affordances and interaction allow to improve the convergence speed of reinforcement learning, and the results also show that the system is robust against wrong instructions that result from errors of the speech recognition system.

Classic pathfinding algorithms face different problems for proper navigation in a dynamically changing environment like in the RoboCup humanoid Soccer league. Working with uncertain and noisy data and missing an effective localization, an intelligent approach was tested. Classic potential fields and a evolutionary trained CTRNN were combined for an improved navigation which includes finding an effective path to the goal, the correct orientation of the robot and obstacle avoidance. This combination was successfully tested in a simplified 2D Simulation.

Unsupervised learning of features or representations has been shown to yield highly competitive results in comparison to hand-crafted features. This can be observed for an increasing number of Machine Learning tasks, most prominently object recognition. However, for related tasks like action recognition in videos, hand-crafted features are still the dominant method of choice. In this talk, I will give an overview of recent research and Neural Network models for unsupervised motion features. Most of these model are extensions of either Restricted Boltzmann Machines or Autoencoders and incorporate higher-order multiplicative interactions. The occurrence of multiplicative units is very common and their necessity will be motivated. Higher-order interactions on the other hand require many additional model parameters and generally result in more difficult training. As an alternative, I will introduce a model that avoids complex higher-order interactions and is much easier to train. This model works by detecting synchrony between temporally separated features using just multiplicative units. Recently, I have successfully generalized this model to arbitrarily sized videos using a weight-sharing / convolutional approach. The presentation will conclude with some open questions regarding the application and design of experiments. Current ideas for future research directions will be given as well.

To enable the localization in the RoboCup Soccer context, the algorithm Rat- SLAM, which does simultaneous localization and mapping, was ported into an existing RoboCup framework. The algorithm was extended to use the data from the existing image processing and to improve thereby the position matching. The use of the distances to the goal posts and of the lines in relation to the robot were examined. It was shown that these additional data can improve the matching of the RatSLAM algorithm.

The efficient recognition of parts of the environment is relevant for a variety of scenarios ranging from lower-level grasping and manipulation to higher-level robotic assistance. In this presentation we will show an approach for scene understanding for assistive robotics based on learning to recognize different objects from RGB-D devices. Using the depth information it is possible to compute descriptors that capture the geometrical relations among the points that constitute an object or extract features from multiple viewpoints. We developed a framework for testing different neural models that receive this depth information as input. Also, we propose an approach using three-dimensional RGB-D information as input to Convolutional Neural Networks. Results reported F1 scores greater than 0.9 for the majority of the objects tested, showing that the adopted approach is efficient as well for classification.

The idea for this gathering is to discuss what a good PhD thesis needs in terms of structure and coherence. For this we want to have a look at some examples for good theses with varying style and scope to provide a good start as well as hints on considerations and pitfalls for PhD students on all levels. Discussions will include corner stones that PhD committees are looking for as well as why a thesis always is individual but also special and knowledge about writing a Master's thesis helps only to some extent.

Three-dimensional (3D) integral imaging allows one to reconstruct a 3D scene, including range information, and provides sectional refocused imaging of 3D objects at different ranges. This paper explores the potential use of 3D passive sensing integral imaging for human gesture recognition tasks from sequences of reconstructed 3D video scenes. As a preliminary testbed, the 3D integral imaging sensing is implemented using an array of cameras with the appropriate algorithms for 3D scene reconstruction. Recognition experiments are performed by acquiring 3D video scenes of multiple hand gestures performed by ten people. We analyze the capability and performance of gesture recognition using 3D integral imaging representations at given distances and compare its performance with the use of standard two-dimensional (2D) single-camera videos. To the best of our knowledge, this is the first report on using 3D integral imaging for human gesture recognition.

Giving interactive feedback, other than well/bad done alone, allows to learn faster in Reinforcement Learning. If we reduce the number of wrong actions the learner takes, e.g. by undoing bad actions, we can reduce the number of steps necessary to learn a task. However, the amount of help needed is not thoroughly researched. This thesis aims to answer this question with two tasks. In both tasks the learner learns to move one arm each to reach for an arbitrary position within its reach. The first task is with a shoulder as joint alone, the second with a shoulder and an elbow. We will discover that the advantage of undoing bad actions manifests itself only in the arm with elbow and shoulder and not in the arm with a shoulder as only joint. We will see that not only did the interactive learner learn faster, it learned to take less steps to reach its target, too. Furthermore, it needed less configuration effort to get good result, while the non-interactive learner did not learn very well in some configurations. If both learners learn well, we need a lot of feedback to show an increase in the behaviour, as well as the learning speed. If the difference between both learners is large, even a probability of 10% of interactive feedback shows an improvement of both properties. With a probability of 60% or more, both properties become more stable than the property of the original learner without interactive feedback.

In the current research on many robots, one goal is to make inexpensive and safe robots for application in domestic environments. Therefore, we have started to develop a humanoid robot called NICU using the existing robot hardware for further extensions to employ in Human-Robot interaction. NICU exhibits the walking capabilities of Humanoid Open Platform Nimbro-OP and the interaction capabilities of i-Cub Head. The existing arms with three degree of freedom have been already transformed into arms with five degree of freedom and also a grip is included. The arm design is not based strictly on any existing arm manipulator, rather it is a new idea implemented by part design and later printing those parts by a 3D printer. The expectation at the end of the research work is to make NICU capable of performing smoothly with new arms, i.e. to perform grasping using neural network techniques and giving a human a friendly look. The talk is mainly about explaining the progress so far, the approaches used and to have discussions about approaches possibly to be considered for future implementations.

In many physical activities such as dancing, sports, and strength training, specific movements have to be learned and perfected. Therefore, it is helpful to receive feedback on how well one is performing a movement and how to improve it. This thesis examines how to give motion feedback automatically by means of learning a model of correct motion and extracting feedback from it. The proposed learning architecture uses a recursive self-organizing neural network model to learn postural and temporal aspects of motion. We present three techniques for feedback extraction from a thus learned model, which give information about the whole body and about individual joints. To evaluate our approach, we captured a body pose dataset of athletes performing compound barbell lifts and dumbbell exercises using a depth-sensing camera. Results show a recall and precision of 0.9 and higher for detection of mistakes in movements on single-subject data.

Natural communication between humans involves hand gestures, which has an impact on research in human-robot interaction. In a real-world scenario, understanding human gestures by a robot is hard due to several challenges like hand segmentation. To recognize hand postures this work proposes a novel convolutional implementation. The model is able to recognize hand postures recorded by a robot camera in real-time, in a real-world application scenario. The proposed model was also evaluated in a benchmark database.

Autonomous agents operating in the real world require a method for representing what they know about the world. Certain recent approaches to representing world knowledge are based in prediction: they describe the world as a set of predictions about what the agent will perceive if it acts in various ways. One predictive method that shows particular promise is the General Value Function (GVF). These GVFs (also called "forecasts") have many benefits, including their simplicity and adaptability, uniformity, flexibility, and semantic interpretability. That is: forecasts can encode, with minimal structure, predictions that can be learned and refined through experience; they can encode small- and large-scale predictions equally well; they can be used without modification for a wide variety of environments and sensorimotor apparatus, and their predictions correspond well to human notions of knowledge. Consequently, forecasts may also more readily capture regularities in an autonomous agent�s sensorimotor stream than other predictive methods. In this talk I will describe recent research that my colleague Tom Schaul and I did to investigate the ability of forecasts to capture such regularities. We generated focused sets of forecasts and measured their capacity for generalization, and then compared the results to a closely related predictive method (PSRs) already shown to have good generalization abilities. Our results indicate quite dramatically that forecasts provide a substantial improvement over PSRs in terms of generalization: the features they encode generalize better to as-yet-unseen parts of the agent's state space and also lead to better value-function approximation in reinforcement-learning agents.

Mark Ring received his Ph.D. from the University of Texas at Austin in 1994, focusing then, as now, on continual learning. He has held research positions at the GMD (now Fraunhofer) Institute in Bonn, Germany, at the University of Alberta in Canada (working closely with Richard Sutton), and at IDSIA in Lugano, Switzerland.

Interactive reinforcement learning constitutes an alternative for improving convergence speed in reinforcement learning methods. In this work, we investigate inter-agent training and present an approach for knowledge transfer in a domestic scenario where a first agent is trained by reinforcement learning and afterwards transfers selected knowledge to a second agent by instructions to achieve more efficient training. we combine this approach with action-space pruning by using knowledge on affordances and show that it significantly improves convergence speed in both classic and interactive reinforcement learning scenarios.

Automatic speech recognition (ASR) technology has been developed to such a level that off-the-shelf dis- tributed speech recognition services are available (free of cost), which allow researchers to integrate speech into their applications with little development effort or expert knowledge leading to better results compared with previously used open-source tools. Often, however, such services do not accept language models or grammars but process free speech from any domain. While results are very good given the enor- mous size of the search space, results frequently contain out-of-domain words or constructs that cannot be un- derstood by subsequent domain-dependent natural lan- guage understanding (NLU) components. We present a versatile post-processing technique based on phonetic distance that integrates domain knowledge with open- domain ASR results, leading to improved ASR perfor- mance. Notably, our technique is able to make use of domain restrictions using various degrees of domain knowledge, ranging from pure vocabulary restrictions via grammars or N-Grams to restrictions of the accept- able utterances. We present results for a variety of cor- pora (mainly from human-robot interaction) where our combined approach significantly outperforms Google ASR as well as a plain open-source ASR solution.

Recently, there has been a significant interest on ambient intelligence for the recognition of human activity. In this context, the classification of human actions and gestures has shown to be a challenging task to accomplish with an artificial system. Despite the prominent use of depth sensors for pose-motion feature estimation, the question how to process this information for efficiently learning human activity is still open. Additionally, in order to operate in natural scenarios, the recognition process must be robust under noisy conditions and exhibit real-time characteristics. In this talk, I will introduce my previous research work on the recognition of full-body actions and hand gestures using self-organizing networks. I will then present a learning framework based on neurobiological evidence for deploying my approach on mobile robot platforms. I will focus on the processing of time-varying data with incremental self-organizing networks, the extension of these architectures for classification tasks, and a pragmantic comparison between feedforwards and recurrent self-organizing architectures for noisy input.

Autonomous robots that can assist humans in situations of daily life have been a long standing vision of robotics, artificial intelligence, and cognitive sciences. A first step towards this goal is to create robots that can learn tasks triggered by visual stimuli from higher level instruction. However, learning techniques have yet to live up to this promise as only few methods manage to scale to high-dimensional manipulator or humanoid robots. In this talk, we investigate a general framework suitable for learning motor skills in robotics including both manipulation of static and dynamic objects that are perceived using vision. The resulting approach relies on a representation of motor skills by parameterized motor primitive policies acting as building blocks of movement generation, and a learned task execution module that transforms these movements into motor commands. We discuss task-appropriate learning approaches for imitation learning, model learning and reinforcement learning for robots with many degrees of freedom that perceive the manipulated objects using robot vision. Empirical evaluations on a several robot systems illustrate the effectiveness and applicability to learning control on an anthropomorphic robot arm. These robot motor skills range from basic visuo-motor skills to playing robot table tennis against a human being and manipulation of various objects.

Biography
Jan Peters is a full professor (W3) for Intelligent Autonomous Systems at the Computer Science Department of the Technische Universitaet Darmstadt and at the same time a senior research scientist and group leader at the Max-Planck Institute for Intelligent Systems, where he heads the interdepartmental Robot Learning Group. Jan Peters has received the Dick Volz Best 2007 US PhD Thesis Runner Up Award, the Robotics: Science & Systems - Early Career Spotlight, the INNS Young Investigator Award, and the IEEE Robotics & Automation Society's Early Career Award. Jan Peters has studied Computer Science, Electrical, Mechanical and Control Engineering at TU Munich and FernUni Hagen in Germany, at the National University of Singapore (NUS) and the University of Southern California (USC). He has received four Master's degrees in these disciplines as well as a Computer Science PhD from USC. Jan Peters has performed research in Germany at DLR, TU Munich and the Max Planck Institute for Biological Cybernetics (in addition to the institutions above), in Japan at the Advanced Telecommunication Research Center (ATR), at USC and at both NUS and Siemens Advanced Engineering in Singapore.

In this thesis we transfer a multisensory reinforcement learning model, developed by Weisswange et al., to an actual robotic platform (an iCub) in a real experimental set-up. We evaluate the issues in doing so and prepare and conduct a localisation experiment for evaluating the performance of the learner with respect to our set-up.

Sound source localisation (SSL) is important for the safety of robots deployed in dynamic environments and for human-robot communication. In this talk I explain how SSL can considerably improve the performance of automatic speech recognition (ASR) with humanoid robots by tracking the position of a speaker. The embodiment of humanoid robots can provide spatial cues that are important for robust SSL. Such spatial cues come from the interaural time (ITD) and level differences (ILD) of sound sensed by the robot. I present a model where auditory spatial cues are represented with spiking neural networks modelling relevant areas in the mammalian auditory pathway. It models the medial and lateral superior olives for the extraction of ITDs and ILDs. Afterwards, both cues are integrated with a model of the inferior colliculus (IC) using Bayesian inference. Finally, a feed-forward neural network is used to classify the IC output. This final layer is not intended to model a particular area of the auditory system. Nevertheless, it is important to handle non-linearities produced by reverberation and ego-noise from the robot. Once the sound source angle is estimated, the humanoid can turn its head to the angle where ASR exhibits the best performance. Interestingly, ASR does not have the best performance at the angle where the ear microphone points directly to the speaker, but at the angle where the ear pinna reflects sound most intensely towards the microphone. To conclude the talk, I will discuss possible extensions of this model to include vision, e.g. for segregating speech at the neural level in addition to the behavioural level. I look forward to your participation in the seminar.

In my current work, I am exploring Restricted Boltzmann Machines (RBM) as building blocks for Fingerspelling (FS) recognition. FS is part of many sign languages and is used by deaf people to manually spell out names and words. My FS dataset is given as a labelled sequence of hand postures, captured and represented densely using the hand pose estimation developed during my Master's Thesis. FS recognition is particularly interesting, because neighbouring postures are not isolated but can form higher level structures bearing complex semantics, like syllables and words. This motivates the use of RBMs, because they can be easily stacked on top of each other to form so called Deep Belief Networks in order to learn more complex structures. RBMs are a subclass of undirected probabilistic graphical models. Despite their origin in probability theory, RBMs have close connections to Artificial Neural Networks (ANN). From this perspective, RBMs are simple feed forward ANNs with an unsupervised Hebbian learning rule. This way, RBMs can be interpreted as mapping observations onto latent units that best explain the input. Certain variants, e.g. the Conditional RBM (CRBM), are also capable of learning sequential data. In this talk I will present the FS recognition task and the FS dataset. Furthermore, (Conditional) RBMs will be introduced as well as some common use-cases and results.

Using Recurrent Neural Networks (RNN) for sequence modeling is one of the fundamental approaches in the area of neurally-inspired computations. Derived from classic dynamical system theory, computations in RNNs trained by Backpropagation Through Time (BPTT) suffer from vanishing error gradients, trapping in local minima and bifurcation. Reservoir Computing (RC) offers an alternative for RNN training by fixing network topology and facilitating learning by simple linear readout of the network weights. In this talk I will give an introduction to RC, especially Echo-State-Networks (ESN), and explain why this approach could be useful for dynamic gesture recognition. In a classification task, I employed ESN as ensembles for 5 commanding gestures. First experiments show promising classification results. In a later stage, I would like to transfer the approach to humanoid robots for application in a Human-Robot-Interaction (HRI) scenario. The steps to achieve this comprise using ESNs also as motor generating network to learn the robot proactively participate in HRI. Also, the multimodal aspect using gestures as visual and language as additional auditive channel are of special interest for my work.

Mapping, localization and navigation are three basic, but important tasks for autonomous robots. They have always been major challenges in robotics. Variations in environment and local conditions force robots to behave more robustly than possible with hard-coded behaviors. Over time many approaches dealing with this problem have been done. Most of them use data collected from odometry or landmarks and a priori generated maps of the environment. During the last decades, ever increasing computational power has led to powerful probabilistic Simultaneous Localization and Mapping (SLAM) algorithms, which deal with real-world sensory data using Kalman filters, particle filters and/or expectation-maximization algorithms (EM). Therefore, SLAM is well investigated, developed and used most frequently nowadays. In the last years, there has been much research in biological motivated mapping and localization approaches. These approaches try to achieve accurate mapping with the use of noisy inaccurate sensors. In 2004, a biologically inspired SLAM approach has been introduced that offers a solution for concurrent mapping, localization and navigation that can be used on real robots. This approach is called RatSLAM. It is based on a computational model of parts of the rodent's hippocampus, which maintains the believed location inside the world. RatSLAM uses landmark sensing and odometric information in combination with a competitive attractor network (CAN) to form a topologically consistent representation with largely Cartesian properties. For now, RatSLAM was used for wheeled robots and for straightforward flying quadrotor platforms. It was never used in combination with humanoids because of their increased constraints. In this presentation an approach is presented that uses RatSLAM as a basis for an adapted mapping and navigation solution for the humanoid robot NAO from Aldebaran Robotics.

Mobility in Wireless Sensor Networks (WSN) are employed today not only as an efficient mechanism to increase the network's lifetime but as an effective cooperative way to perform difficult tasks ranging from coverage to exploration in complex environment. There are several research issues in both planning and control of robots carrying out missions in synergy with WSNs. Not only there are relevant optimization problems that need be tackled, but also questions related to the inherent stochastic nature of real robots, real sensor nodes and real environments. The objetive of this seminar is to present planning and decision making strategies that can be serve as inspiration to develop solutions that model the rational behaviour of a team of robots acting cooperatively.

The dissemination of robotic technologies for domestic purposes was the driving force for our Master project "Human-Robot-Interaction" (HRI). The goal of our project was to focus on how HRI-related important topics like vision, navigation and behaviour features the purpose of robotics in a domestic environment. A simple task for humans, yet difficult to implement on robots, was to navigate from the KT kitchen to different other office rooms searching for used cups. The project simplified the object recognition approach by implementing an interacting senario between the robot and the users in the room where for the used cups was asked for. In this presentation, we will present our algorithms in more detail and give insight into the performance of our experiments, which had been carried out in the KT environment. The talk is followed by a live-demonstration.

The well-known machine learning algorithm Reinforcement Learning uses Q-values, which spread the information about which action to take from the goal state to the surrounding states. Thereby, over several trials, the information is back-propagated. Experiments with rats or human beings, on the other hand, have shown that decision-making can be seen as forward thinking, respectively thinking in creating plans and executing them. The thesis investigated different machine learning algorithms, both backward and forward oriented, and tested their behavior in a dynamic environment, in which the goal condition changes over time. The Q-value-based algorithms all suffered from deprecated Q-values after each goal change. A planning algorithm, which only learns a model of the environment and does not use Q-values, performed best in small environments, but suffered from an exponentially growing size of possible paths to check. So on the one hand, planning algorithms that use Q-values can reduce the number of planning paths to check effectively if the Q-values are useful. On the other hand, they can compensate for the deprecated Q-values to some extent.

Cloud-based automated speech recognition (ASR) systems employ large acoustic models which lead to good performance in a wide domain like web searches or dictations. In restricted domains (like ones using a constrained vocabulary), these systems often do not perform as well as local ASR systems because they are not adjustable to that domain due to API restrictions. (Users are able to pass audio data to the system, but no further configurations are possible like passing lan- guage models.) We investigated the option of post-processing the results of cloud-based ASR systems instead of employing adjustable local ASR systems. The post-processing heuristics are based on phonetic distances and different matching techniques and differ in the extent of applied domain knowledge by ranging from vocabularies over statistical language models and grammars to a fix list of sentences. To be able to compare our approach with related approaches we chose corpora from the human robot interaction (HRI) domain and a well-known standardized corpus. The results of our experiments show that most of our heuristics significantly outperformed the Google ASR system and a local open-source ASR system. We also observed that providing more domain knowledge to the post-processing system leads to better performance.

Every robotic arm has to reach in order to grasp an object and reinforcement learning allows the agent to learn this task. To date, multiple approaches have been shown to work but it may take many steps until the learner reaches the target reliably. Interactive machine learning allows a human or other external, more knowing agent to help the learner along. It may reduce the number of steps to reach the target for the first time by "nudging" it towards a better position or giving more informative feedback. In my thesis I intend to explore how these additional influences can be used to augment an existing reinforcement learning algorithm and if or how much they increase the learning performance. I plan to test this augmented algorithm in a reaching task with a simple 2 dimensional 2 DoF arm.

In many physical activities such as dancing, sports, and strength training, specific movements have to be learned and perfected. Therefore, it is helpful to receive feedback on how well one is performing a movement and how to improve it. This thesis aims to develop a learning framework for providing feedback on the correctness of human body movements. The system will use a feature extraction module for capturing relevant properties of body motion. A recurrent self-organizing neural network architecture will be then used to learn a set of correct movements. At detection time, the system will estimate a measurement on the correctness of the test motion sequence and provide qualitative feedback on how to fix mistakes in the dynamics of body postures. The framework will be evaluated on datasets containing both correct and wrong body motions from depth map video streams.

Natural language is the cognitive capability that clearly distinguishes humans from other living beings but is by far not yet understood. In contrast to other cognitive functions or methods to obtain precise information how language is processed and - more important how - language has developed in the brain are very limited. Traditional generativists theories on language processing are often problematic in explaining 'the how', because they neglect the involvement of embodiment and situated context, i.e. the close relation of language processing and proprioception. Recent constructivists theories are often limited to grounded items on single word level, not capturing the complex temporal dynamics. In recent research we aimed at bridging this gap in terms of searching for the crucial characteristics of a neural architecture that allows for the emergence of language. With a MTRNN-based architecture that combines the processing of complex phonetic utterances with visual experience we showed that the MTRNN inherits some important characteristics in terms of connectivity, plasticity, and temporal abstraction. In the upcoming talk I want to follow up on these studies and report on the recent refinements and extensions of the architecture. In the new multimodal architecture that also reflects the cell assemblies theory we aim at taking motor movements into account and show that the modelled characteristics allow for the abstractions of concepts from visual and proprioceptive perception as well as the generalised verbalisation of concepts in natural language.

It is possible to think of daily life examples where sound source localisation improves the understanding of language in noisy environments, such as when elderly people with auditory deficits turn their heads to 'face' the speaker with their ear instead of their eyes. Following this heuristic, we explored the impact of such turning behaviour for improving speech recognition using the iCub humanoid head. More specifically, we assessed the feasibility and impact of this approach by measuring 1) the time it takes the robot to face a sound source from different starting angles, and 2) whether a specific angle between the sound source and the robot's face improves speech recognition. After initially obtaining promising results, a second round of experiments displayed oscillations in the system's performance. However, further inspection in the mechanism used for extracting sound signals revealed a clear advantage in the use of sound source localisation for improving speech recognition.

Over the last more-than-two years, I have been theorizing about multi-sensory integration. I have developed a model of multi-sensory integration in the superior colliculus (SC) and shown that it both performs near-optimally and reproduces certain aspects of the neurophysiology of the SC and the behavior of SC-using animals... in computer simulations. It is high time that I applied this model in a physical experiment. In the first part of this talk, I will speak about the goals and motivation of my PHD project and give a description of the model I have developed. This will be followed by a discussion of the biological phenomena this model reproduces and a blue-skies outlook on some of the next steps in modeling I want to take. The second part of the talk will be concerned with the currently pressing issue of designing a robotic experiment for my model. To give you an intuition of the conceptual problems I am facing, I will first give you my thoughts on what I think an experiment can do for me in principle. After that, I will explain some technicalities of the infrastructure available for the experiment, the kind of output I want to generate, and criteria for the input. The talk will end with an invitation to the audience to give their thoughts and suggestions.

This bachelor thesis investigated two different kinds of robot simulation programs for comparison, namely Webots and the open-source simulator V-Rep. The former is well-established for robotic simulations but expensive, the latter is more cost-effective and thus interesting for future usage in robotic projects. An important aspect in the simulation of robots is the simulation of humanoid robots, with NAO (a specific humanoid robot manufactured by French robotics company Aldebaran Robotics) taking a special interest in this field due to its widespread use. Although the V-Rep environment does provide an existing model for the NAO robot, that model is not controllable through NAOQi. In this regard, the question arises whether V-Rep can be adapted to control the NAO just as easily as Webots can. The NAO runs on a software called Naoqi. The simulator Webots offers already a robot controller that uses the functionality of Naoqi, making it easier for the programmer to specifically tell the robot what the user wants him to do. In V-Rep, this function is not available. After extensive research, it was found out how exactly the NAO is able to be controlled via V-Rep. This research consisted of examining various control methods, followed by an explicit comparison between V-Rep and Webots in terms of their complexity, implementation, as well as similarities and differences In the end, it can be concluded that V-Rep represents a suitable alternative to Webots. Although not all of the functionality was tested, it was shown that V-Rep has, in some parts, even advantages over Webots, for example the choice of physics engines and the possibility to run V-Rep much faster than Webots.

The versatility and adaptability of behaviour demonstrated in nature is usually ascribed to the large neuro-mechanical redundancy in the body of organisms. The resulting high-dimensionality however complicates our understanding of the principles underlying coordinated movement control and learning; the so-called "Curse of Dimensionality" in motor control. Elucidating the principles by which the Central Nervous System (CNS) alleviates this curse in achieving motor coordination is vital for both understanding natural intelligence as well as for developing novel control techniques for complex robots. In this talk, a mitigation strategy based on dimensionality reduction in control is presented, for the facilitation of learning real-time movement coordination. In the first part, the dimensionality problem and conventional techniques for planning and control in robotic systems are reviewed along with the neuroscientific hypotheses of optimal motor control, developmental skill acquisition, and muscle synergies. In the second part, a control-theoretic framework is presented for a synthetic analysis of reduced dimensional behaviour in systems. The framework is used for analysing the different factors facilitating dimensionality reduction in systems, i.e. (i) natural dynamics, (ii) output (task-space relevance) and, (iii) input (modularisation of the control). On the basis of various simulation experiments, a principle of exploiting reduced dimensionality is proposed for the design and control of systems that must autonomously acquire motor skills. The implications of this principle are then discussed for the neuroscientific theories of motor coordination and development and for the design and control of high-dimensional artificial systems such as biomimetic robots.

Short Bio: Naveen Kuppuswamy holds a Master of Science (MS) in the field of Electrical Engineering from the Robot Intelligence Lab of the Korea Advanced Institute of Sience and Technology (KAIST). He successfully defended his PhD-thesis concerning goal-oriented motor control in the Artificial Intelligence Lab within the group of Prof. Dr. Rolf Pfeifer (University of Zurich, Switzerland). He is affiliated in recent research projects like RobotDoc and Amarsi.
For more information: Naveen Kuppuswamy

In this talk we will present our work according to HaarNN-Ensembles. HaarNN-Ensembles use Haar-feature-like pattern as input for Hebb-learning Neural Networks combined with the ensemble learning algorithm called 'Adaboost'. We consider three different models of HaarNN: First, the Bayes-like stable pattern model which combines the stable states of an Hopfield Neural Network with the Bayes method [1]. The second method is in analog to the first, but uses the cluster algorithm K-means for grouping the input. The third approach groups the input pattern for the different classes through an moving average and has online learning ability.

[1] Meins, N., Magg, S., Wermter, S. Neural Hopfield-ensemble for multi-class head pose detection. Proceedings of International Joint Conference on Neural Networks (IJCNN). pp. 1327-1334, Dallas, 2013.

Die hier vorgestellte Arbeit beschäftigt sich mit der Navigation von mobilen Robotern und gliedert sich in zwei Arbeitsschritte. Zunächst wurde das von W. Yan et al. [1] vorgestellte Framework in der Simulationsumgebung Webots implementiert und damit für Simulationen nutzbar gemacht. Damit ist es möglich, Roboter in Webots mit der von Yan vorgestellten Navigation zu betreiben. Im zweiten Schritt wurden diese Roboter verschiedenen, aufeinander aufbauenden Testszenarien ausgesetzt, um dort ihre Fähigkeit zur Navigation auf die Probe zu stellen. Diese Tests wurden für diese Arbeit dokumentiert und ausgewertet, um schließlich Rückschlüsse zu ziehen, inwiefern die Navigation mit verschiedenen Umgebungen, aber auch mit unterschiedlichen Robotern funktioniert.

[1] Yan, W., Weber, C., Wermter, S. A Neural Approach for Robot Navigation based on Cognitive Map Learning. 2012 International Joint Conference on Neural Networks (IJCNN2012), pp. 1146-1153, Brisbane, Australia, 2012.

Predictive mechanisms exist hierarchically in the cortical areas, which compensate delays via prediction so that the cortical areas encode present, not past, events. To realize such compensation in the context of artificial neural networks and cognitive robotics, we suggest to use recurrent connections to create an internal memory to store the previous dynamics. In the first experiment, we emphasized that a short-term memory should exist within the mirror neuron system to assist with the action understanding and prediction. A recurrent network with parametric biases was thus applied to realize this system function in which the parametric bias units recognize and encode walking patterns from another robot. Furthermore, in the second experiment, we focused on the encoding of information in the primary visual cortex. Specifically, the convergence of ventral and dorsal pathways with predictive functions may give rise to the understanding of object affordance and object manipulation and its control. To model this, we designed a recurrent predictive network with a horizontal product where the information of object feature and object movement becomes successfully separated in its two hidden layers. Since the dorsal pathway also allows for motor-relevant representations, in the third experiment, we also developed such recurrent connections for sensory latency compensation which also support smoother and faster behaviours in sensorimotor integration tasks. We expanded the use of recurrent connections in the sensorimotor system, particularly in the sensory prediction part, so that the recurrent connections could compensate the delay in the sensory percepts. A continuous actor-critic automaton (CACLA) was also applied in the generation of smooth behaviours corresponding to the predictive sensory percepts. In the last experiment, we combined the first and second experiments and proposed that the learning of the visual pathways also leads to the conceptualization of visual information. This was realized by recurrent networks with a horizontal product too, which was extended with parametric bias units. We exemplified this model through a robot passively observing an object to learn its features and movements and meanwhile the pre-symbolic representation was self-organized in the parametric units.

Self-defensive and survival neural circuits belong to the most important and essential capabilities of any organism. They are believed to impose the ground rules for more complex and motivated behaviour. Although many of these innate responses are hard-coded in the brain, they are not sufficient for the organisms' survival. They have to adapt to new and unexpected situations within their lifetime and thereby be able to interact effectively with their environment. A key component on the lifetime adaptation is the formation of associations/memories between environmental predictors and relevant events, which mainly rely on punishment and reward learning. We believe that a deeper understanding of innate and learned defensive mechanisms could also be helpful in developing future robot generations, making them more adaptable and robust. In this research project, we are studying and developing neurocomputational self-protective mechanisms for a humanoid service robot such as association of appetitive stimuli and more importantly association of aversive or noxious stimuli. Our first experiment addressed reward-driven learning where a robot is trained to seek for appetitive stimuli. Here a reinforcement learning (SARSA) algorithm was optimized to learn in a real-world scenario and manoeuvre a humanoid robot towards a charging station. On a second experiment we study the role of noxious stimuli on the formation of anticipatory behaviour. This experiment based on the vast literature on Pavlovian and instrumental conditioning and how environmental cues can be used to anticipate negative outcomes. A hybrid approach using an echo state network (ESN) and a dopamine modulated Pavlovian conditioning was used to anticipate painful signal based on auditory cues. Finally, our current research focuses on emulated pain signals due to its important role in attracting attention and modulating decision making and action. Pain is an unpleasant sensory experience associated with actual or potential tissue damage and thus a strong indicator for behaviour readjustment. However maladaptive behaviours may arise leading to persistent avoidance behaviour and ironically to more pain and long-term disability. Particularly, we are studying the boundaries between adaptive and maladaptive withdrawal behaviours in a motor learning task. Specifically, a continuous actor-critic automaton (CACLA) algorithm optimized for real-world learning is used to teach a humanoid robot to reach for a moving target.

The communication task is one of the most important regarding the Human-Robot Interaction (HRI). It is the main form of cooperation between humans and robots, and thus several studies have been made focusing the natural language interpretation. One of the main points of the human-robot communication is the joint attention using gestures and speech. This kind of communication leads to scenarios when a human can express things like: " Robot, get me that tool" or "Give it to me" always pointing to some object. This problem has a multi-modal temporal nature and thus is still a complex pattern to be recognized. My current PhD work proposes the use of deep architectures allied with computer vision and speech recognition techniques to solve this multi-modal problem. In this talk the motivation and the background for this particular problem will be presented, as the project plan to walk through this problem using the deep architectures advantages. An application scenario and the methodology of this work will be discussed as well as the expected results.

In the last 5 years there has been a relevant increase for research on ambient intelligence for the recognition of human activities and behavioral patterns in indoor environments. The recent introduction of low- cost range sensors using structured light technologies (e.g. Microsoft Kinect and ASUS Xtion) allows to efficiently estimate object distances from the observed scene and obtain a mapping in terms of both projective and real world coordinates, overcoming a set of historical limitations imposed by most color cameras. Not surprisingly, this novel sensory trend led to an increasing number of vision-based applications using depth maps instead of, or in combination with, RGB pixel matrices. The use of depth information has shown to increase robustness under varying light conditions and significantly reduce computational effort for demanding visual tasks such as the tracking of moving human targets. However, despite the development of new sensor technologies and an exponential growth in computational power, the human brain continues to outperform vision-based applications for object segmentation, tracking and behavior analysis in terms of performance and accuracy. One of the major reasons for this flaw is the disregard of the dynamics of the human visual system for encoding and processing visual stimuli to solve those visual tasks that a human observer can achieve with little effort. It is therefore of great importance to study convenient motion representations and hierarchical neural architectures that better take advantage of processed depth information. In this seminar I present the state of the art in learning systems for human action recognition using depth images and analyse the use of biologically inspired architectures to develop applications for real time behavior analysis. For this purpose, I describe a set of strategies derived from neurocognitive evidence to extract and process relevant motion features from an observed scene and use them as input for a hierarchy of self-organizing networks, e.g. SOM and other extensions. I conclude with future directions on human activity recognition and the introduction of mobile robot platforms to extend depth sensors for active tracking.

Face detection has experienced a growing interest in the last decade with the rise of intelligent real-time systems. Applications range from video surveillance to mobile apps and tagging in social media. We have explored the performance of neural network ensembles performing face detection by applying perceptron learning to the Viola Jones Object Detection Framework. We have trained single and multi-layered perceptrons as weak classifiers, which use Haar-like feature templates as inputs. These networks are grouped into ensembles by AdaBoost. Our focus is on the binary classification problem, although we also achieve successful multiclass head pose detection (three viewing directions) with multi-layered perceptrons. In this thesis we evaluate the impact of boosting on the neural networks, as well as the use of different types of feature templates. Our results show that neural network ensembles can yield a high detection rate with a low number of false positives.

Humans learn actions on different ways, one of them is researching our environment and trying repeatedly up to achieve a desired aim. Learning may be supported by e.g. a person advising another one in the process of using a coffee machine the first time. Reinforcement Learning has demonstrated being a useful learning approach, but, as recent work has shown, lacks often performance. This work looks for improving learning speed of a robot within human agents interaction. It is proposed to use a domestic scenario, where a human supports a robot in cleaning tasks of a plane surface (e.g. a table). In this talk first ideas and scenarios will be presented. In addition, the proposal invites discussion about using interactive reinforcement learning for action prediction and neuro-biologically motivated representations.

Webots is a state of the art robot simulator used by WTM but with the high price tag isn't there something else we could use? I will explore a recently introduced simulation environment called V-REP to simulate robots and try to find out what's different to Webots in general and whether possible improvements are worth the effort to switch to V-REP. Since WTM already has an extensive body of work done for Webots, easy transfer of existing modules for the NAO robot to V-REP is essential to justify a move towards the new environment. I will try to explore and compare several ways of merging existing code into the V-REP environment and the outcome of this project will be (one or more) modules to interface NAOqi and therefore previous work with a simulated NAO in V-REP. Additional analysis on simulation performance will follow to complete the analysis of possible advantages and disadvantages.

This thesis deals with the experiment of training different SOM set ups for the task of grasping in three dimensions with a robot arm. The training and recognition is based on data provided by one, respectively two cameras. Three different SOM set ups were trained and evaluated. One set up was a single SOM trained with the data of one camera(1x2D), the other setup was one SOM trained with data of two cameras (1x4D) and the last set up consisted of two SOM, each trained with the data of a different camera(2x2D). The results suggests that given the context of the experiment the 2x2D is the superior configuration and that the set ups that involved both cameras simultaneously are significantly more precise than the set up trained with data from only one camera.

We present a hierarchical SOM-based architecture for the detection of novel behavior in indoor environments. The system can unsupervisedly learn normal activity and then report novel behavioral patterns as abnormal. The learning stage is based on the clustering of motion with self-organizing maps. With this approach, no domain-specific knowledge on normal actions is required. During the tracking stage, we extract human motion properties expressed in terms of multi-dimensional flow vectors. From this representation, three classes of motion descriptors are encoded: trajectories, body features and directions. During the training phase, SOM networks are responsible for learning a specific class of descriptors. For a more accurate clustering of motion, we detect and remove outliers from the training data. At detection time, we propose a hybrid neural-statistical method for 3D posture recognition in real time. New observations are tested for novelty and reported if they deviate from the learned behavior. Experiments were performed in two different tracking scenarios with fixed and mobile depth sensor. In order to exhibit the validity of the proposed methodology, several experimental setups and the evaluation of obtained results are presented.

Multi-class object detection is perhaps the most important task for many computer vision systems and mobile robots. In this work we will show that Hopfield Neural Network (HNN) ensembles can successfully detect and classify objects from several classes by taking advantage of head-pose estimation. The single HNNs are using pixel sums of Haar-like features as input, resulting in HNNs with a small number of neurons. An advantage of using these in ensembles is their compact form. Although it was shown that such HNNs can only memorise few patterns, by utilising a naive-Bayes mechanism we were able to exploit the multi-class ability of single HNNs within an ensemble. In this work we report successful head pose classification, which presents a 4-class problem (3 poses + negatives). Results show that successful classification can be achieved with small training sets and ensembles, making this approach an interesting choice for online learning and robotics.

When confronting binaural sound source localisation (SSL) algorithms with different environments and robotic platforms, there is an increasing need for non-linear integration methods of spatial cues. Based on interaural time and level differences, we can compare the performance of neural vs. statistical SSL methods. For this purpose, we implemented a testing architecture with three degrees of freedom, i.e. a different combination of a) representation of binaural cues, b) clustering and c) classification algorithms. The heuristic for the selection of methods is the same at each degree of freedom: to compare the impact of traditional statistical techniques versus machine learning algorithms with different degrees of biological inspiration. The overall performance is evaluated in the analysis of each system, including the accuracy of its output, training time and adequateness for life-long learning. The results support the use of hybrid systems consisting different kinds of artificial neural networks, as they present an effective compromise between the characteristics evaluated.

Hand pose estimation is the task of deriving a hand's articulation from sensory input. Traditional approaches to this challenging problem rely on either costly hardware or measure directly at the user's hand (e.g. data-gloves). This thesis presents a CV-based method, which captures the complete kinematic state of the hand. Recent research indicates that the Kinect is a viable device to implement marker-less pose estimation, thus enabling unobtrusive and natural HCI applications. These approaches formulate pose estimation as an optimization problem: a high dimensional hypothesis space is constructed from a hand model, in which particle swarms search for the best solution. We propose various changes to this novel approach. For preprocessing, we do not rely on skin color, but exploit the properties of depth images and recognize hands by their shape. Our hand model is extended to include anatomical constraints of hand motion by applying a PCA. This also allows us to treat pose estimation as a problem with variable dimensionality. The most important benefit becomes visible once our PCA-enhanced model is combined with biased particle swarms: accuracy and performance of pose estimation are improved significantly. Several experiments were conducted as part of this thesis. These give proof to the superior properties of our pose estimation.

The superior colliculus (SC), or optic tectum (OT) as it is called in non-mammals, is a prime example of natural multi-sensory integration. A brain region found in all vertebrates, it combines auditory, visual, and tactile information to locate events in space, and generates motor output resulting in orienting behavior. The talk will give a brief introduction to/refresher of SC physiology, on the behavioral and neurological levels, motivate the merit of modeling the SC from the perspective of a computer scientist, and then discuss why and how self-organization and statistics are good tools for the task.

Primates can learn complex sequences that can be represented in the form of categories, and even more complex hierarchical structures such as language. The prefrontal cortex is involved in the representation of the context of such sequential structures, and the striatum permits the association of this context with appropriate behaviours. In my PhD work, the Reservoir Computing paradigm is used: a recurrent neural network with fixed random connections models the prefrontal cortex, and the read-out layer (i.e. output layer) models the striatum. The model was applied to language syntactic processing, especially thematic role assignment: given a sentence this corresponds to answer the question "Who did what to whom?". The model processes categories (i.e. abstractions) of sentences which are called "grammatical constructions". The model is able to (1) process correctly the majority of the grammatical constructions that were not learned, demonstrating generalization capabilities, and (2) to make online predictions while processing a grammatical construction. Moreover, we observed that when the model processes less frequent constructions an important shift in output predictions occurs. It is proposed that a significant modification of predictions in a short period of time could be related to Evoked-Related Potentials (ERP) like the P600 which typically occurs when unusual sentences are processed. Finally, to show the ability of the model to deal with a real-world application, the model was applied in the framework of human-robot interaction for both sentence comprehension and production. The sentence production model was obtained by "reversing" the sentence comprehension model: it processes meanings as input and produces sequence of words in output.

Short Bio: Xavier Hinaut received an M.S. in Computer Science from the University of Technology of Compiègne (France) in 2008 and an M.S. in Cognitive Science from the Ecole Pratique des Hautes Etudes (France) in 2009. In early 2013, he obtained a PhD in Computational Neuroscience at the INSERM Stem Cell and Brain Research Institute in Lyon (France) under the supervision of Peter Ford Dominey. His research explores brain mechanisms of complex sequence processing, in particular on language syntactic comprehension. He is interested in using Recurrent Neural Networks to model these processes and investigating how such models can be used in developmental robotics.

Self-defensive and survival neural circuits belong to the most important and essential capabilities of any organism. They are believed to impose the ground rules for more complex and motivated behaviour (Sternson, 2013). Pain signals have a key role in these processes attracting attention and motivating decision and actions. Although pain signals are quickly linked to their role in reactive behaviour, they also play a key role in learning, for instance leaning to predict pain enables agents to avoid injuries. Moreover we believe that pain signals may also speed up learning and lead to smoother and safer behaviour. To test this hypothesis we have design an object reaching experiment with a Nao humanoid robot. Here we use reinforcement learning (Continuous Actor-Critic Learning Automaton, a.k.a. CACLA) to test the impact of pain-like signal on learning. Learning from scratch approach is used to compare learning speed and subjective aesthetic robot's pose when reaching for an object.

How the human brain understands natural language and what we can learn for intelligent systems is open research. Recently, researchers claimed that language is embodied in most - if not all - sensory and sensorimotor modalities and that the brain's architecture favours the emergence of language. In this talk I will report our recent study about the characteristics of such an architecture and propose a model based on the Multiple Timescale Recurrent Neural Network, extended by embodied visual perception. The study shows that such an architecture can learn the meaning of utterances with respect to visual perception and that it can produce verbal utterances that correctly describe previously unknown scenes. In addition I would like to open a discussions about integrating visual sensations for multiple objects in the scene over time to investigate the characteristics of an architecture that is supposed to reflect relations between objects in natural language.

As a fundamental research topic, autonomous indoor robot navigation is challenging in robotics. Although many models for map-building and planning exist, it is difficult to integrate in the real world due to the high amount of noise and complexity. Following our previous work [1], we present a neural cognitive model for environment mapping and robot navigation based on learning spatial structure by observing the persons' movement using a ceiling-mounted camera. A robot can plan and navigate to any given position in the room based on this map, and adapt it based on interaction with the obstacles. In addition, consistent with the suggestion of [2] that a correspondence between the notions of visual appearance and spatial location representations exists, salient visual features are learned and stored in the map during navigation. This anchoring of visual features in the map space enables the robot to find and navigate to a target object by showing an image of it. We implemented this model on a humanoid robot and tests are conducted in a home-like environment. Results of experiments are evaluated and we conclude our work with an outlook for the future development.

[1] W. Yan, C. Weber, and S. Wermter, "A neural approach for robot navigation based on cognitive map learning." in Proceedings in the 2012 International Joint Conference on Neural Networks (IJCNN2012). Brisbane, Australian: IEEE, June 2012, pp. 1146-1153.
[2] M. J. Farah, K. M. Hammond, D. N. Levine and R. Calvanio, "Visual and spatial mental imagery: Dissociable systems of representation." Cognitive Psychology, Volume 20, Issue 4, October 1988, pp 439-462.

Text categorization is the task of discovering the category or class text documents belongs to, or in other words spotting the correct topic for text documents. While many machine learning schemes for building automatic classifiers exist today, these are typically resource demanding and do not always achieve the best results when given the whole contents of the documents. A popular solution to these problems is called feature selection. The features (e.g. terms) in a document collection are given weights based on a simple scheme, and then ranked by these weights. Next, each document is represented using only the top ranked features, typically only a few percent of the features. The classifier is then built in considerably less time, and might even improve accuracy. In this work we implemented different feature selection algorithms, create feature vectors of different lengths and use them together with different classification algorithms. The quality of the classification is used to examine the interactions between the algorithms. We found that the concrete selection algorithm is less important for the resulting quality of the classification than the length of the feature vector. To compute our experiments faster, we used a parallel implementation for the feature selection and the classification task.

Kevin Köser is a senior researcher in the Computer Vision and Geometry Lab at the Swiss Federal Insitute of Technology (ETH) Zurich. His research interests lie in 3D computer vision problems, such as shape and motion extraction from video, visual geolocalization, registration and tracking and recently involved novel feature representations and geometric relations as well as efficient robust matching schemes. He has worked in several national and international collaborations on 3D reconstruction, registration and recognition with industrial partners such as the BBC, Bosch, Keystone, Leica Geosystems, Navteq, Nokia, Swiss Television SF, Testo, Volkswagen and XSens. Dr. Köser obtained his PhD (Dr.-Ing.) from the University of Kiel for his thesis "Geometric Estimation with Local Affine Frames and Free-form Surfaces" (dissertation award 2009) and his research was awarded several prizes, most notably the DAGM 2011 Main Prize (German Association for Pattern Recognition). He also serves as a member of the program commitee for the major international conferences on computer vision and in robotics and teaches the classes "3D Photography" and the "Computer Vision Lab" at ETH Zurich.

The purpose of the bachelor thesis was the evolution of artificial neural networks to develop locomotion for the Darwin robot. The DARwIn-OP, in the future referred to as Darwin, is a 50cm tall humanoid robot which is used, amongst others, in the RoboCup for robot soccer. The main problem in robot soccer is a robust and fast locomotion. Since a humanoid robot is a very complex system, it is difficult to handcraft a robust walking algorithm. Furthermore it needs to be adjusted by hand if the floor or the weight distribution of the robot itself is changed. An approach to automatically develop a walking algorithm is biologically in- spired evolution in which a gradual improvement of individual solutions can be achieved over many generations. The advantage of evolution is that the problem itself does not need to be solved by hand, but only needs a so-called fitness function, which rates the quality of a solution. In addition, the execution of the required experiments can be parallelized very well. But evolution has also certain difficulties which must be overcome. For example, tens of thousands of experiments need to be performed in order to find a good solution in a complex search space. We have developed a system, which exploits the concurrency offered by evo- lution and performs the experiments in the Webots simulator on many different computers in parallel, therefore finding solutions in a reasonable amount of time. It uses a realistic model of the Darwin, which makes a transfer of a good solution to the real Darwin realistic. This would be beneficial for further support of the RoboCup soccer team, which can achieve even better results in future championships. The focus of this work is on oscillating pattern generation within the artificial neural network (ANN) and by external sources, as well as the impact of neurons in the hidden layer of the ANN. The experiments have shown, that an ANN is able to generate a pattern without the use of a central pattern generator. Furthermore the results indicate that at least 4 neurons in the hidden layer have to be present to evolve a locomotion.

Face detection is an active research area com- prising the fields of computer vision, machine learning and intelligent robotics. However, this area is still challenging due to many problems arising from image processing and the further steps necessary for the detection process. In this work we focus on Hopfield Neural Network (HNN) and ensemble learning. It extends our recent work by two components: the simultaneous usage of different image representations and combinations as well as variations in the training procedure. Using the HNN within an ensemble achieves high detection rates but shows no increase in false detection rates, as is commonly the case. We present our experimental setup and investigate the robustness of our architecture. Our results indicate, that with the presented methods the face detection system is flexible regarding varying environmental conditions, leading to a higher robustness.

The talk looks at situated collaboration between humans and robots, for example in complex situations like disaster response. This collaborative context gives rise to several issues. One, we need to start from the assumption of *asymmetric agency*: Humans and robots experience and understand the world differently. The asymmetry implies that symbols cannot be considered abstract types, embodying an objective truth. Instead, different agents employ different types in building up understanding, or more precisely they construct subjective judgments as proofs for why a particular type can be applied to some experience. This then gives rise to another issue, namely how these different actors could then arrive at some mutually shared understanding or ''common ground.'' We see this as the need to align judgments, and formally construct this as an alignment between (abductive) proofs over multi-agent beliefs and intentions. This places grounding meaning in context in a new light: Grounding meaning, between actors, is the alignment of judgments against subjective experience, within a social, situated context. The intentional aspects of social dynamics play thereby just as much a role as do beliefs. In a collaborative activity, meaning becomes socially and situationally construed referential content, with a semi-persistent nature that is subject both to social dynamics (why the meaning is construed and used) and environment dynamics (what the meaning is construed for). This requires an actor, particularly a robot, to be explicitly aware of these social dynamics and its own role(s) in it. The talk captures this through the notion of social sentience.

As the brain attempts to compensate the latency of its sensorimotor cycle, a biologically-inspired learning model of predictive sensorimotor integration is proposed to compensate that latency in a robotic system. In this model, an Elman network is employed for sensory prediction and a Continuous Actor-Critic Learning Automaton (CACLA) trained for smooth action generation. For a robot approaching experiment, this architecture also improves the smoothness of the robot's trajectory and therefore brings a faster and more accurate approaching behavior. Experimental results prove that this novel architecture improves the approaching success-rate and decreases the approaching time.

In order to make robot vision closer to human vision, we have to further progress brain-inspired vision system technology, consisting of three different subunits of "recognition," "eye movement control" and "memory." We here present current development of the recognition subunit. We focus on improving the Elastic Graph Matching (EGM) for object detection. The EGM is often regarded to be one of neural models for visual object recognition that are successfully applied into industries, but still has unsolved problems for the real-time processing. To solve such problems, we suggest different scale similarity map and then develop the real time EGM-based object detection, comparable to the existing computer vision technology. We will present briefly a binocular movement control model, which will be integrated with the EGM-based object detection for realizing an object tacking in the near future.

Vision-based human action tracking and novel behavior detection represent a cornerstone of data-driven scene analysis. I will present a learning framework for real-time novel behavior detection from depth maps video sequences. Based on neurocognitive evidence, the framework prioritizes the extraction of spatio-temporal properties, such as location and direction of motion. Unlike other related work discussed in the literature, my framework does not require domain- specific prior knowledge on performed actions. Spatio-temporal patterns are processed to obtain suitable representations for unsupervised clustering of the extracted data. In the field of novelty detection, self-organizing maps (SOM) have shown to be a prominent method for learning patterns from training data. Techniques and experimental results on SOM-based novelty detection will be presented and discussed.

Neural fields are a popular method of neural information processing, because of their intuitive topographic representation of neural activity. However, their centre-surround competitive connectivity is usually hand-designed or trained with the wanted activity pattern applied directly to all units. I will present an Elman network in which the hidden layer learns to become like a neural field, while direct feedback is given only to the output layer. The example network has a one-dimensional "chain" of neurons on each, input, hidden and output layer. A static/moving hill of neural activation on the input layer shall be reconstructed/predicted by the network on the output layer. Topographic weight initialisation and certain constraints on weights and activations make the network learn the desired centre-surround connectivity on the hidden layer, and a topographic static/moving activity hill emerges there, too. The network also learns to predict an activity hill that can be either static or move in either of two directions. Then the connectivity becomes more complicated and the network's mechanisms deserve further investigation.

Estimating a detailed human hand configuration in 3D space is a challenging task but has diverse interesting applications in the field of Human-Computer Interaction. These include natural, intuitive user interfaces, learning by demonstration and speech-replacing gestures, either static or dynamic. The availability of cheap depth sensors in the recent years (Kinect, Xtion) offers new opportunities to approach vision-based gesture recognition. One possibility for a gesture recognition system is to provide a full hand description for correct pose estimation. Here, a discrete set of possible poses is defined for searching a continuous space for optimal solutions. One approach to this task was proposed by Oikonomidis et al. (2012). They reformulate the problem as an optimization problem, which is then solved using standard Particle Swarm Optimization. But the proposed method suffers from two drawbacks: (1) it uses only basic preprocessing, resulting in uncertainties during the early stages and (2) a high computational cost effectively preventing real-time applications. Thus, the goal of this Master thesis is to improve upon their work. Early image preprocessing will be replaced by previous work done at WTM. To achieve real-time performance, certain key areas in their theoretical foundations have been identified, which will be revised. Based on a working hand pose estimation, a system for static gesture recognition will be implemented. This system will be used to evaluate the advantages and limitations of the refined methods.

Simultaneouslocalization and mapping (SLAM) has always been one major challenge in robotics. Variations in environment and local conditions forcerobots to behave more robustly than possible with hard-coded behaviors. Over time many different approaches dealing with this problem have been done. Most of them use data collected from odometryor landmarks and pre-generated maps of the environment. Increased computational power has led to powerful SLAM algorithms, which deal with real-world sensory data using Kalman filters and/or expectation-maximization (EM) algorithms. Therefore SLAM is well investigated and developed. In 2004 a new approach in SLAM has been introduced: ratSLAM. It is based on a computational model of specific parts of the rat hippocampus, which maintains the rodent's believed location in the world. In this model, landmark sensing and odometric information are used by a competitive attractor network to form a topologically consistent representation with largely Cartesian properties. For now ratSLAM has only been used and tested on wheeled robots. In this talk I will give an overview of the current state-of-the-art. Furthermore, problems that appear when bringing ratSLAM to the next level by integration in a humanoid robot are pointed out and possible solutions are given.

Modelling locomotion is a very interesting but challenging research area in robotics, especially by using a bio-inspired neuronal controllers. Recently, the approaches to modelling motor abilities on robots have developed a lot by systematically applying CPGs (central pattern generators) to different sorts of robotic platforms including both quadrupedal and bipedal robots. In this talk, in the first part, a relatively new CPG architecture (an integration of oscillatory recurrent neural network and feed-forward network ) is proposed to represent the parts controlling the rhythmic movement like crawling and walking. The biological grounds for this architecture is also introduced. In the second part, an interesting learning architecture, CPG-ACTOR-CRTIC, is introduced to involve the learning mechanisms and proffers potential meanings for exploring motor learning capabilities of infants.

There are different approaches for Automatic Speech Recognition systems to map the audio input into a string of phones which will then be processed further. Most systems use Hidden Markov Models to build up a model of each phone. Lately the use of neural networks of different kinds has grown. The advantage of these networks is the greater robustness against noise. A drawback is the problem of learning long time dependencies. The error between the said phone and the recognised phone is used to train the networks. For long lasting phones this error vanish or blow up. With the proposition of a new type of artificial neurons, called 'Long Short Term Memory', the problem of vanishing errors is solved and there were studies on neural networks for Automatic Speech Recognition containing this new type of neuron. For speech recognition on robots there is the problem of strong noise coming from the robots fans or motors. Based on the greater noise robustness of neural networks we examine the advantages of Long Short Term Memory neural networks compared to Hidden Markov Model on an Automatic Speech Recognition task of audio data, recorded on a humanoid robot of type iCub. We show that the predominant noise of the robot prevents any correct recognition results on systems based on Hidden Markov Models, as well as on systems based on neural networks.

In this final project presentation, a hybrid learning system is introduced as the result of the human-robot interaction (HRI) project. The system uses a Nao robot for interaction, which can be controlled by speech. It is able to learn presented objects and recognize them afterwards. Furthermore, the system can differentiate between learned objects and point at the object asked for. A hybrid system is used, which consists of region-of-interest (ROI) extraction, different feature extractors and a neural network for each feature extractor. The outputs of each network are used for ensemble learning. In a performance test, the system recognizes 100 percent of the learned objects. The system consists of many loosely coupled modules which are connected via a robot operating system (ROS) to be able to distribute the computations to several machines. The project presentation includes a talk and an interactive demonstration and provides room for interesting discussions.

In my talk, I am going to present a robotic connectionist model of the contribution of gesture to learning to count. The model is based on a recurrent artificial neural network architecture, connected with the iCub humanoid robot. I compare counting performance of the model in two conditions: without and with gesture, showing that the proprioceptive signal connected with gesture indeed carries information which may be exploited when learning to count. I also compare the behaviour of the model with that of human children, based on current experimental psychology literature. Finally, I present the prospects for future work, analysing the ways in which the model can be extended.

The presentation focuses on the trajectory planning task for object grasping with reference to the NAO robot. The intermediate trajectory for movement from a specific initial arm configuration to a final one is modeled using Bezier curves and the parameters of the curve are optimally determined using energy minimization criteria. The optimization is performed using a bio-inspired evolutionary algorithm, namely Differential Evolution. The talk will focus on the algorithmic aspects of the trajectory planner along with discussions on the feasibility for online implementation on the NAO robot.

The success of an ensemble of classifiers depends on the diversity of the underlying features. If a classifier can address more different aspects of the analyzed objects, this allows to improve an ensemble. In this talk I will present an ensemble using as classifier members a Hopfield Neural Network (HNN) that uses Haar-like features as an input template. I will analyse the HNN as the only classifier type and also combine it with threshold classifiers to a hybrid neural ensemble, so that the resulting ensemble contains --as members-- threshold and neural classifiers. This ensemble architecture is evaluated for the domain of face detection. I will show that a HNN that uses summed pixel intensities as input for the classification has the ability to improve the performance of an ensemble.

In dieser Arbeit wird ein mehrstufiges Objekterkennungssystem am Beispiel von Gesichtern vorgestellt, die von einer humanoiden Roboterplattform aufgenommen wurden. Vier Stufen kennzeichnen das Verfahren, ein Farbfilter, eine Normalverteilung, eine "Support Vector Machine" (SVM) und eine einfache Methode, die sich der Hauptkomponentenanalyse und einem Abstandsmaß bedient, das bestimmt, ob es sich um ein Gesicht handelt oder nicht. Dieses Verfahren beinhaltet nur eine Personendetektion. Zu einem Gesicht wird ein Merkmalsvektor erzeugt, der verschiedenen Eigenschaften, wie dem Bild als solchem, Histogramme, Gradienten und weitere inne hat. In den ersten beiden Stufen des Verfahrens findet mittels Diskriminanzanalyse eine Klassifikation in die Kategorien: Gesicht, Nicht-Gesicht oder eventuelles Gesicht statt. Die SVM erlernt eine maximal trennende Hyperebene aus den Trainingsdaten und entscheidet bei einem Testbild auf welcher Seite der Hyperebene die Projektion landet, um so die Klassenzugehörigkeit zu bestimmen. Abschließend wird von den übriggebliebenen Kandidaten mit Hilfe von "Eigengesichtern" (eigenfaces) ein Abstandsmaß berechnet, um zu bestimmen, ob es sich um ein Gesicht handelt. Um die Resultate zu verbessern, soll die Trainingsmenge verändert werden. Den Fragen, wie wichtig ist die Wahl der Trainingsmenge, wie kann die Trainingsmenge verändert werden um bessere Ergebnisse zu erzielen, welche Vergrößerungsmethoden bringen etwas, welche Rolle spielt die Methodik bei der Erstellung, wird im Rahmen dieser Untersuchung nachgegangen. Methoden zur Veränderung der Trainingsmenge, wie Streckung und Stauchung, werden angewandt, um eine Verbesserung der Trainingsdaten zu erzeugen. Untersuchungen haben gezeigt, dass die getesteten Methoden, Verbesserungen bei der Klassifikation erbracht haben.

In this workshop, I will present an overview on automatic speech recognition (ASR). The main components like language models, dictionaries, grammars, search managers and acoustic models will be briefly explained. Then I will show how ASR is implemented in three different open- source frameworks and how they can be used: First, I will present the "Julius" engine, how to install it on a Linux-based systems and how the different components are defined by a configuration. Second, I will introduce the the RWTH automatic speech recognition (RASR) system. Third, I will show how the CMU Sphinx tool-kit can be used on the NAO robot and how some shell scripts can be used to generate the different components. Finally, I will have a closer look at the acoustic model training with sphinxtrain.

Sources:
http://www-i6.informatik.rwth-aachen.de/rwth-asr/
http://cmusphinx.sourceforge.net/

I will present the performance results of a spiking neural network for binaural sound source localisation (SSL) using the Nao robot. In order to explore the effect of the robot's embodiment on the sound signal, the cues used for SSL were the interaural time (ITD) and level (ILD) differences. As biological plausibility was pursued in our implementation, ITDs and ILDs were extracted with models of the medial superior olive (MSO) and the lateral superior olive (LSO). The influence of the robot's body on the sound signal was quantified through a statistical analysis of the MSO and LSO response of sounds produced at different azimuth angles. These descriptive values proved essential to achieve accurate SSL. Finally, the outputs from the MSO and LSO were integrated in a model of the inferior colliculus (IC). The connection weights between the MSO and LSO neurons to the IC neurons were estimated using Bayesian inference. This inference process allowed the algorithm to perform robustly under high levels of ego-noise.

Keywords: Bio-mimetic binaural sound source localisation, Spiking neural networks, Bayesian inference.

We present an algorithm based on the self-organizing map (SOM) which models multi-sensory integration as realized by the superior colliculus (SC). Our algorithm differs from other algorithms for multi-sensory integration in that it learns mappings between modalities' coordinate systems, it learns their respective reliabilities for different points in space, and uses mappings and reliabilities to perform cue integration. It does this in only one learning phase without supervision and such that calculations and data structures are local to individual neurons. Our simulations indicate that our algorithm can learn near-optimal integration of input from noisy sensory modalities.

With the focus on CTRNN as a method for grounding utterances in natural language in omni-modal experiences, recent findings in training a state-of the art CTRNN will be presented and a novel idea, in terms of extending such a network to take into account all available modalities, will be discussed. The talk will be divided into two parts:
1) Recent research has revealed that hierarchical linguistic structures can emerge in a recurrent neural network with a sufficient number of delayed context layers. As a representative of this type of network the Multiple Timescale Recurrent Neural Network (MTRNN) has been proposed for recognising and generating known as well as unknown linguistic utterances. However the training of utterances performed in other approaches demands a high training effort. To overcome this issue, we analyse the training demands and propose a robust mechanism for adaptive learning rates and internal states to speed up the training process substantially. In addition we compare the generalisation of the network for the adaptive mechanism as well as the standard fixed learning rates finding at least equal capabilities.
2) The MTRNN revealed interesting capabilities but also limitations: To generalise and therefore generate untrained utterances, the network needs to "recognise" the initial state value for some context units. In addition no integration of multi- or omni-modal experience have been accomplished so far. As part of current and future work an extended MTRNN will be discussed that overcomes these issues. For this Oberseminar also some preliminary results of generalising utterances will be outlined.

Classical recommender systems solve an information filtering task. They suggest a data object, typically a text document describing some product or service, that is likely to be relevant to the user based upon her or his previous choices. A geographic recommender system recommends items from a library of geo-referenced objects. In multi-object recommendation, collections of items are suggested which should consist of somehow similar exemplars but, at the same time, must show variability. A geographic multi-object recommender suggests, for instance, a list of cities to visit or a slide show of images illustrating a certain place. The talk gives an introduction into the research field and presents the approach taken by our Tripost multi-object recommender for selecting a small set of geo-referenced images of a touristic location. Motivated by this research, we ask how different conceptualizations of a city can be identified in web-based image collections. The method is illustrated by analyzing image data set from the cities of Amsterdam, Bamberg, Cardiff, and Dublin. In addition, the talk presents a recently collected data set of GPS tracks and geo-referenced photographs taken from visitors of the Old Town of Bamberg. Our data suggests that differences in the frequency of spatial choices need to be taken into account when building place models for recommender systems.

Gesture recognition provides a natural interface for Human-Robot-Interaction. Until now, several important methods have been exploited in order to correctly recognise and classify gestures, for instance, Neural Networks and Hidden Markov Models (HMM). However, most techniques rely on additional technical devices to ease data capturing and classification process. Vision-based gesture recognition in contrast is most user-comfortable, but also the most challenging approach to the recognition problem. In this talk we will motivate usage of statistical models like Conditional Random Fields (CRF)[1]. A CRF is a derivative of the well-known Markov Networks, which have been extensively used in computer vision problems like image segmentation and object recognition. As CRFs are probabilistic graphical models, it can handle missing or erroneous data by reasoning under uncertainty, thus further motivating application in the field of Artificial Intelligence. Recent studies showed good results compared to state-of-the-art models like HMM, so we will extend this formalism to the gesture recognition problem concerning commanding gestures, which will be further used in a human-robot-scenario. Future work on graphical models in terms of different topologies to model sub-gestures and so to enhance the prediction and therefore recognition accuracy will be discussed.

[1] Klinger, R. & Tomanek, K. Classical Probabilistic Models and Conditional Random Fields, Technical Report, Department of Computer Science, Dortmund University of Technology, 2007

According to the theory of two parallel visual pathways[1], the 'dorsal pathway' encodes spatial information, invariant of stimulus-specific properties, while the 'ventral pathway' encodes object feature identity, invariant of positions and sizes. In this talk, I will propose a novel biologically-inspired neural architecture constitutes of a horizontal product in a multi-layer network in order to acquire this property of the visual system. Experiments show that, learned in an unsupervised manner, two sets of hidden units represent cells in the ventral and dorsal pathways, respectively. Moreover, the fast responding dorsal-like cells have become direction selective while the slow responding ventral-like cells encode 'object identity', which are analogous to the recording of different populations of complex cells in V1 [2].

[1] M. Mishkin, L. Ungerleider, and K. Macko, "Object vision and spatial vision: Two cortical pathways," Trends in Neurosciences, vol. 6, pp. 414-417, 1983.
[2] N. Priebe, S. Lisberger, and J. Movshon. Tuning for spatiotemporal frequency and speed in directionally selective neurons of macaque striate cortex. Journal of Neuroscience, 26(11):2941-2950, 2006.

In this talk we present a neural network architecture for a robot learning new navigation behavior by observing a human?s movement in a room. While indoor robot navigation is challenging due to the high complexity of real environments and the possible dynamic changes in a room, a human can explore a room easily without any collisions. We therefore propose a neural network that builds up a neural memory for spatial representations and path planning using a person?s movements as observed from a ceiling-mounted camera. Based on the human?s motion, the robot learns a map that is used for path planning and motor-action codings. We evaluate our model with a detailed case study and show that the robot navigates effectively.

RoboCup@Home is the largest domestic service robot benchmark and arguably also the most complex for any fully autonomous robot to participate in. The benchmark is set up as a competition, where the robots can do a series of tests in an apartment. The tests become more and more complex over time and usually have the form of a story, such as that the robot has to behave as a butler, or that somebody lost an item and the robot has to find it. In the benchmarks we test for all the aspects that we deem necessary to create a social robot. Since nobody exactly knows what it takes to create a social robot, even our benchmarks are being put under scientific scrutiny in order to improve on them. We avoid the trap of moving towards some local optimum of robotics bodies and minds by changing the benchmarks every two years. In order to keep track of the improvements we have developed methodologies for statistical benchmarking. This is needed because the robots change, but also because we under-specify many aspects of the competition, both for the tests and for the environments, since the real world is also under-specified. We aim to create a general purpose social robot that can be put to use in many situations. The robot has to be able to behave itself under highly uncertain and dynamic situations. Therefor we take the robots outside of the apartment and into the shopping mall. We started with a very successful 'shopping' test. This year we plan to take the robots for a walk and take the elevator to get to the shopping mall. Since the robot has to be fully autonomous, there is a large focus on artificial intelligence. We increased the focus on AI by introducing the first test ever that focuses on the understanding of the robots with regards to their environment. This test is so unstructured that we have a computer program that generates the problems that the robot has to solve, of course within certain boundaries.

Symbol systems provide techniques to emulate processes of the human mind, such as abstract reasoning, natural language and problem solving, in artificial intelligent agents. To purposefully employ these techniques in artificial agents, however, these symbols need to be grounded, i.e. the agent must be able to connect a symbol to its real world referent and vice versa. Marocco et al. (2010) [1] proposed a method that uses neural networks to ground symbols in the sensorimotoric experience of a humanoid robot. After training, the networks were capable to identify three different situations and utter a corresponding proto-word. We tried to employ the proposed method on a different humanoid robot, NAO. In the process, we identified and investigated the parameters that in uence the stability of the proposed method and the quality of its results.

[1] Davide Marocco, Angelo Cangelosi, Kerstin Fischer, and Tony Belpaeme. Grounding action words in the sensorimotor interaction with the world: experiments with a simulated icub humanoid robot. Frontiers in Neurorobotics, 4:1-15, 2010.

Many scientific works have shown the usefulness of artificial neuronal network based robot controllers. The choice for simple robots has been object approach, for example approximating the position towards geometrical line patterns, while for humanoid robots it has been the bipedal walking problem. As the latter is still a challenge, the approach of finding a neuronal network based controller by means of an evolutionary algorithm was topic of the bachelor thesis. Although none of the found controllers reached expectations, the process itself seemed to work properly. The results of the project including the approach and an analysis of known issues will be presented.

In this work, we present a neurocomputational model for auditory-cue fear acquisition. Computational fear conditioning has experienced a growing interest over the last few years, because of its possible implications for clinical treatment of human anxiety disorders. Fear learning is a simple and robust learning paradigm that involves sensory and motor aspects [1]. We argue that a deeper study of the mechanisms underlying fear circuits in the brain will contribute not only to a better conceptual understanding of neural fear processing in general but also to the development of safer robots. Our specific model is based on LeDoux's dual-route hypothesis of fear and also dopamine modulated Pavlovian conditioning. Our reservoir approach is capable of learning the temporal relationship between auditory sensory cues and an aversive or appetitive stimulus. The model was tested as a neural network simulation but it was designed to be used with minor modifications on a robotic platform. Applications of this learning mechanism may include artificial self-protective systems, autonomous battery recharging [2] or detection of emergency situations.

[1] H.-C. Pape and D. Pare, "Plastic synaptic networks of the amygdala for the acquisition, expression, and extinction of conditioned fear, Physiological reviews, vol. 90, no. 2, pp. 419-463, 2010.
[2] N. Navarro, C. Weber, and S. Wermter, "Real-world reinforcement learning for autonomous humanoid robot charging in a home environment," in Proceedings of the Annual Conference Towards Autonomous Robotic Systems (TAROS), ser. Lecture Notes in Computer Science, R. Groß, L. Alboul, C. Melhuish, M. Witkowski, T. Prescott, and J. Penders, Eds., vol. 6856. Berlin, Heidelberg: Springer-Verlag, 2011, pp. 231-240.

Foraging tasks are commonly used in homogeneous or heterogeneous swarm experiments since the principles behind foraging are similar to other tasks (e.g. collecting, harvesting, search, etc). In this talk I present work done as part of my PhD thesis, in which I've extended the results of a previous robot foraging experiment by Labella. Labella had shown that a group of agents, adapting their activity according to their individual success, has a consistently higher efficiency in terms of collected objects per time spent outside their nest. By analysing the foraging behaviour of homogeneous and heterogeneous groups using his adaptation algorithm, I've been able to identify the reasons responsible for the efficiency increase and show that - even for my very abstract multi-agent scenario - efficiency does not necessarily increase when adapting agent activity to individual success rates.

Um eine solide Interaktion mit einem menschlichen Benutzer möglich zu machen, muss jeder Roboter in Zukunft ein umfangreiches auditorisches System bereitstellen. Neben der Spracherkennung ist dabei auch die Geräuschquellenlokalisation ein notwendiger Bestandteil. Diese Arbeit zeigt die Realisierung eines Lokalisierungsverfahrens in der horizontalen Ebene, mit einem humanoidem Roboter unter Nutzung von Cross Correlation und Künstlichen Neuronalen Netzen. Der Roboter soll 24 verschiedene Richtungen am vollen 360 Grad Kreis, mit einem jeweiligen Abstand von 15 Grad, unterscheiden. Dies gelingt dem vorgestellten Verfahren mit einer Genauigkeit von bis zu 92%. Dabei treten die fehlerhaften Klassifikationen hauptsächlich zwischen benachbarten Richtungen auf.

In principle biological systems have to reduce complexity of possible states to obtain an efficient functional structure. These states may be different in nature. One may consider the protein space of a molecule, interactions between molecules in a cell, interactions in a neural network or interactions between biological agents. During the evolutionary process a functional structure has to develop a sufficient fitness or it will end. A genetically determined protein may appear to evolve slowly. However, the neurophysiological system of an agent has to adapt to biological relevant characteristics of the information flow in order to be able to react immediately on demands of the environment. Thus, it is a challenging question, what these different levels of building a functional adaptive structure have in common. An attempt will be put forward about combining schema theory with the approach of reaction-diffusion systems. Experimental results of movement control are evaluated on the basis of such a model.
Reaction-diffusion systems provide a rich variety of activity patterns. Different types will be discussed. How do biological patterns emerge and how does the behaviour of an agent lead to a schema. The protein space can be considered to be like a picture, each possible protein and each pixel to represent a point in a high dimensional space, respectively. Such a picture may be disturbed or noisy. However, if besides others a certain picture is stored in an association matrix a recurrent attractor net may reconstruct the original picture. However, reaching a stable attractor point, that is a static picture, would be boring. On the basis of specific association matrices a recurrent system can be constructed which provides a sequence of consecutive states. This leads to the hypothesis that complex movements are based on such consecutive states of motion primitives.
From a modeling point of view, it should be possible to develop a neural network which allows for implementing a schema, i. e. a motor engram, which interacts with feedback control in an iterative manner. Neural population coding is supposed to represent a suitable modeling approach. Considering two interacting neural layers of a reaction-diffusion system they may generate an attractor state and a moving ‘bubble’, respectively. The bubble running along a motor engram activates a set of muscles and causes the arm to move towards a target. Modeling results are evaluated against the background of experimental data obtained from hand movements. They support the hypothesis of a motor schema. It determines the trajectory of intercepting a moving target. It comes up, that in conjunction with the motor schema the transition probability of the bubble elicits a new motion primitive and its direction during the next step. The Gestalt of the movement trajectory is determined by a schema which is adapted to the requirements of the task.

Moving away from engineering maxim of "simple is beautiful" towards biological statement of "complexity is not a problem". Automatic classification or prediction model building directly from data with a minimal or no human supervision is already absolutely crucial in order to stay competitive and maximally exploit the data in quickly changing business environments. However, the methodology for ensuring that created predictive models and systems are as good as possible should be in place before using them with confidence. No human supervision in model building or operation also implies that one should use powerful enough techniques which: a) can learn the data to any degree of accuracy and b) be able to adapt to changing environments. In order to address these challenging issues and following various inspirations coming from biology coupled with current engineering practices, we propose a major departure from the standard ways of building adaptive, intelligent predictive systems and moving somewhat away from the engineering maxim of "simple is beautiful" to biological statement of "complexity is not a problem" by utilising the biological metaphors of redundant but complementary pathways, interconnected cyclic processes, models that can be created as well as destroyed in easy way, batteries of sensors in form of pools of complementary approaches, hierarchical organisation of constantly optimised and adaptable components. The results of extensive testing for many different benchmark problems and various snapshots of interesting results covering the last decade of our research into predictive modelling will be shown throughout the presentation and a number of challenging real world problems will be discussed in the context of the main goals of a major EU INFER project which we currently coordinate.

Various optimality principles have been proposed to explain the characteristics of coordinated eye and head movements during visual orienting behavior. However, the neural substrate of the underlying computations has been left unspecified. At the same time, researchers have suggested several neural models to underly the generation of saccades, but these do not include online learning as a mechanism of optimization. Here, we suggest an open-loop neural controller with a local adaptation mechanism that minimizes a proposed cost function. Simulations show that the characteristics of coordinated eye and head movements generated by this model match the experimental data in many aspects, including the relationship between amplitude, duration and peak velocity in head-restrained conditions, and the relative contribution of eye and head to the total gaze shift in head-free conditions. Our model brings together an optimality principle, a neural architecture, and a local learning mechanism into a unified control scheme for coordinated eye and head movements. The model may well generalise to further open-loop controlled movements, such as fast reaching movements.

Eine Grundschule hat bei einem Wettbewerb 30 Tischtennisschläger gewonnen. 2 weitere spendet ein Lehrer. Zu jedem Schläger gehören 3 Bälle. Alles soll an die 8 Klassen verteilt werden. Wie viele Schläger und Bälle bekommt jede Klasse ?''. Auch wenn die durch eine solche Aufgabe gegebenen Rechnungen durch Computeralgebra Programme lösbar sind, so gibt es nur wenige Programme, welche auch Textaufgaben bearbeiten können. In dem Vortrag wird zunächst ein sehr kurzer Überblick über diese bisherigen Systeme gegeben. Da die meisten neueren Programme auf einem ähnlichen Ansatz basieren wird dieser, sowie einige damit verbundene Probleme, knapp dargestellt. Anschließend wird dann ein, im Rahmen einer Masterarbeit, entwickeltes System vorgestellt, welches in der Lage ist, Textaufgaben zu lösen. Hierbei wird insbesondere eine neue Form der Wissensrepräsentation eingeführt, um die Struktur einer Textaufgabe darzustellen, mit welcher sich mehrere Probleme vorangegangener Systeme lösen lassen. Des Weiteren wird auf die Transformation von Texten in ein entsprechendes Modell eingegangen. Abschließend wird eine kleine Evaluierung des Systems sowie weitere Ausblicke besprochen.

This short talk will provide an introduction to parallel programming and the parallel programming API OpenCL. Furthermore some experiences in using OpenCL for implementing a training algorithm including the possible speedups will be reviewed.

In this interactive talk, a wrap up of the Cognitive Neuroscience Robotics workshop and the (largest) International Conference on Intelligent RObots and Systems 2011 will be given and discussed. Topics include recent developement in humanoid and assistive robotics and emerging novel research directions. Particulary interesting papers related to research topics in the group will be provided afterwards.

In this talk a Bachelor project of implementing a model of robot-world interaction [1] on a NAO will be proposed: The robot is supposed to push an object - like a ball or a box - and observe the objects' action. At the same time it perceives a word and should learn to connect this word with the observed action (grounding). These grounded names will be learned with a recurrent neural network and should enable the robot to classify an observation correctly later. The main goal of this project is to get insight into the possibilities and limitations of this model, applied on a real NAO robot.

[1] Davide Marocco, Angelo Cangelosi, Kerstin Fischer, and Tony Belpaeme. Grounding action words in the sensorimotor interaction with the world: Experiments with a simulated iCub humanoid robot. Frontiers in Neurorobotics,4(7), 2010.

In this talk I would like to present my work on evolution of body structures using Webots. After a short overview about my psychologically motivated research in general, I will provide a closer look at what I did and how evolution was incorporated into Webots. A 3D body plan similar to stick figures was evolved together with a neural network controller, fully within the 3D Webots environment using Matlab for the controllers. I will discuss problems I've encountered and how those were (or could be) solved, which might be useful to everyone who will attempt to do similar simulations with dynamic changes in robot structure and parameters within Webots.

Mathematical cognition is a branch of cognitive science that aims at understanding various aspects of numerical processing in the brain. The field has been receiving an increasing attention during the past two decades, and due to the contributions from such fields as experimental psychology, neuroscience and computational modelling, substantial progress has been made. Another tool that could be used for the benefit of mathematical cognition studies is robotic modelling - which has a principal advantage over traditional computational modelling methods in a sense that it inherently takes into account the aspect of embodiment. This is in line with the growing amount of evidence that many aspects of mathematical cognition (including the number concept itself, despite it's pure and abstract appeal), are deeply rooted in our very basic interactions with the environment. In the work presented here we look at certain manifestations of an intrinsic connection between representations of number and space in the brain, namely the size and distance effects, SNARC effect and Posner-SNARC effect. Building on previous purely computational modelling approaches, we formulated a robotic model of such interactions. This enabled us to achieve two goals. First, we obtained a model that is much less arbitrary than previous computational models. Second, we showed how certain characteristic patterns of connections, crucial for reproducing aforementioned behavioural phenomena, can emerge as a result of a simple developmental process due to the robot morphology and environmental correlations.

Using visual features for robotic navigation in natural and changing environments remains to be an unsolved problem. This paper describes the use of nearest neighbor algorithm on salient visual features for indoor mapping. The mapping procedure consists on defining an arbitrary number of vector fields -or paths- in a domestic environment, pointing towards a set of target locations. From these specified paths, the robot constantly estimates its most probable location, adjusts its orientation and moves forward a predefined distance, in order to converge to a target location. A SURF (Speeded Up Robust Features) implementation was used for object detection and scene recognition. For every location in a path, the perceived scene was represented with the set of detected objects, and their particular geometric properties relative to the robot's body. A first experiment confirmed that the deviation from the trained paths -due to odometric error- decreased as the number of trained intermediate-locations increased. A second experiment showed promising results when testing the effectiveness of nearest neighbor search for estimating the robot location. Finally in a third experiment, a satisfactory convergence rate of the robot towards predefined target locations, was observed under real world circumstances. Keywords: Topological robot-mapping, visual object-recognition, nearest neighbor, vector-field navigation.

In most activities of daily living, related tasks are encountered over and over again. Countless times, we flip light switches, insert keys in locks, pour coffee, brush our teeth, etc. To exploit this regularity, humans reuse existing motor skills for recurring tasks. For robots, using a set of motor skills also drastically reduces the search space for control, and thus facilitates learning. Therefore, my main research goal is to leverage the advantages of the skill-centric approach to achieve autonomous robots that operate flexibly, robustly and safely in human environments. The specific topics I will consider in this presentation are: * Representing robotic skills with Dynamic Movement Primitives. * Optimizing the robustness, efficiency and safety of skills through model-free reinforcement learning in high-dimensional spaces. * Grounding symbolic knowledge about skills in observed behavior, and exploiting this knowledge to generate sequences of skills for more complex tasks. * Learning task-relevant perceptual features in face model fitting applications, and my current ideas on how this can/should be used in the context of skill-based robotics. Since assistive robots in clinical and household environments must be able to manipulate their environment, my main application domain for evaluating these methods is (mobile) manipulation.

After GasNets were introduced by Husbands et. al. in 1998, they were used for evolution of controllers in many different tasks - from pattern generator tasks to quadrupedal walking. The findings in these experiments were that GasNets can be evolved faster than the same controller type without gaseous connections. Other studies also compared GasNet controllers to other controllers using continuous time recurrent (CTRNN) or probabilistic (PNN) neural networks. In terms of evolvability (length of evolutionary run until a successful controller is found or also robustness of evolved controllers), GasNets always outperformed other network types or the GasNets without gas in these experiments. In this talk, I am going to introduce GasNets and present work done for my MSc thesis. In this research one possible reason for the evolutionary advantage - the propensity to form pattern generator sub-networks - of GasNets was investigated. It was shown that they were outperformed by CTRNNs on a simple shape discrimination task, in which those sub-networks should not lead to an advantage. Different variants of CTRNNs and GasNets were investigated to find the cause of the performance differences and additional results on pattern generator tasks were presented at the end which finally led to the rejection of the initial hypothesis.

In this talk we propose a Hopfield neural network (HNN) working on the basis of Haar feature for visual object detection and more specificially for frontal faces. We will compare our HNN model with the Threshold classifier and use of the Haar feature as Viola and Jones has proposed it. The motivation is to have more benefit from the feature itself with the ontop HNN Classifier instead of the threshold classifier. As a result we try to figure out inhowfar the HNN contributes to a more robust and precise ensemble. We will train our ensembles for comparison with the ensemble algorithm AdaBoost. We will show our results with frontal faces datasets.

The presentation presents a camera-based identification and localization of objects. In this regard, a unique 3d surface reconstruction is performed for each object, in which object-specific features are assigned to each surface point. The reconstruction is based on different perspective camera views. In these images features are sought which can be reliably detected in other views. This allows finding feature correspondences between images, thereby enabling the 3d reconstruction. The so-called 3d feature model (surface points + features) is used to identify an object in an arbitrary image and to determine its spatial position and orientation. Furthermore, an articulated robot performs mobilization activities. With the known spatial pose gripping and moving operations can be performed on objects.

In the last decades, gesture recognition systems based on different input devices, for instance data-gloves, have been developed. Recent applications employ acceleration data derived from according wrist-watches or the Wiimote. Although these approaches have their advantages, they hinder the ease and naturalness for the user in the Human-Computer-Interaction (HCI). In contrast, vision-based systems provide an intuitive interface. However, major limitations are due to typical low-level vision problems like segmentation, illumination changes and self-occlusion, to name a few. Additionally, most algorithms do not meet realtime constraints necessary for online-processing of the gestures. This talk will give an introduction to gesture recognition in the context of HCI. Therefore, the most important developments will be introduced and an overview of the individual processing stages will be given. A major focus will be on the significant methods employed so far in this research area. Finally, a neuroscientifically inspired model will be presented, serving both for implementation and extensions. The theoretical implications will be discussed with the audience. The practical issues will be presented in a following talk (part II).

This talk will be subdivided into two parts. After an introduction of the original Recurrent Neural Network with Parametric Bias (RNNPB) architecture an improved version (iRNNPB), with respect to stability, speed of convergence and accuracy is presented. The applied changes that lead to this boost in performance are elucidated and examples are given showing the intriguing capabilities of this recurrent neural network, including storage and retrieval, recognition and generalization of time series.
In the second part the iRNNPB is applied in a real world robot scenario for action-driven classification of object categories. It will be shown how the architecture handles real world multi-modal sensorimotor data acquired from a humanoid robot and how it compares to canonical methods used for classification. Furthermore, the generalization potential concerning object attributes, its robustness against noise and how this can be used for de-noising of sensory channels is demonstrated.

In the last decade the debate has been intensified if humans have a amodel system for language and higher cognition or if everything we think or talk about is grounded in the modalities of our sensory, sensori-motor or motor systems. Additionally we received neuroscientific evidence for the second view and valuable models of how language is likely to be distributed over the brain. Models, we have seen so far were restricted to single words on the one hand and to words which were closely related to specific motor actions and regions on the other.
From the point of view of a computer scientist I will wrap up in short these research and the open chances to contribute. Furthermore I will present some new ideas about models, which capture the distribution of words in the brain that are related to different audio-video input and some higher concepts like traits or interrelationships of objects. I want to shape these ideas in a PhD project proposal and discuss the implications for application on cognitive robots.

Im Bereich der heutigen Serviceroboter ergibt sich die Problematik, dass der gewöhnliche Anwender sich selten mit speziellen Eingabemethoden auskennt. Da bei zwischenmenschlicher Kommunikation Sprache als Hauptmedium genutzt wird, liegt es nahe auch die Mensch-Roboter-Interaktion (MRI) durch natürliche Sprache zu realisieren. Sei es als Serviceroboter in der Gastronomie, zu Hause oder als Rettungsroboter in Kriesengebieten - in allen Bereichen ist es für den Roboter wichtig, dass er nicht nur erfährt, welche Aufgabe ihm mittels Sprache gestellt wird, sondern auch woher die Information kommt. Beispielsweise soll ein Serviceroboter im Haushalt mit einfachen Mikrofonen festellen können woher ein Hilferuf kommt, um darauf entsprechend zu reagieren. Wenn der Roboter die Geräuschquelle gefunden hat, können andere Teilsysteme des Roboters weitergehende Informationen sammeln (Bsp. Videobilder). Durch welche Methoden eine Geräuschquelle im Raum georten werden kann ist hierbei die Hauptfragestellung, die bearbeitet werden soll. Die Motivation diese Fragestellung zu bearbeiten kommt daher, dass zwar schon die Grundlagen für einige Verfahren in den neunziger Jahren bekannt waren, den Robotern es jedoch an Rechenleistung fehlte. Heutzutage ist die Technik soweit, das auch autonome Roboter in der Lage sind die relativ intensiven Berechnungen durchzuführen. Dies spannt ein attraktives Anwendungsgebiet auf.

This talk is split into two parts. In the first part, features of the new NAO laser head are firstly introduced. With the help of this laser head, it is possible to accurately scan indoor environment and create the map from it. Secondly, some previous works about SLAM based on laser sensor will be presented. From the past decades, many researchers attempt to solve the problem of simultaneous localization and mapping, commonly called SLAM.Many papers use different ways to implement or optimize SLAM methods, but most of them concentrate a small part of SLAM. With representative methods of each part of SLAM, the motivation of this talk is to help ourselves understand SLAM better and to explore future research possibilities arising from SLAM.

Our objective is to research into the foundations, representations and applications of intelligent systems. This research in computer science is focused on Knowledge Technology and Knowledge Management motivated by natural systems, e.g. biology, cognition and neuroscience. We want to study and exploit nature-inspired, in particular hybrid neural and symbolic representations, in order to build next-generation adaptive knowledge-based systems, learning ambient intelligence systems, multimodal neural agents, self organising knowledge discovery systems and human robot interaction systems. Applications include medical reasoning under uncertainty, intelligent retrieval systems, interactive systems, adaptive engineering, data/text mining systems, cognitive and neuroscience-inspired robots, speech/language systems, intelligent web agents and hybrid techniques for medical diagnosis.

Recent neuroscientific evidence suggests that the human amygdala mediates performance in many reward-based decision-making tasks. This group of nuclei located deep within the medial temporal lobes seems to represent the affective/emotional valence, i.e. state value, of a situation, necessary for coordinating physiological, behavioral and cognitive responses. These findings directly link to computational neural reinforcement learning (RL) techniques that are well established within our group. In general, computational models dealing with emotional circuits, especially fear, are rare. Here we suggest a novel model comprising microcircuits for fear acquisition, processing and extinction. The model aims at filling the computational modeling gap between neurophysiological methods and neuroimaging studies. In addition to a computational study we plan to implement the proposed neural architecture on a real-world robot leading to a demonstrator capable of predicting aversive and appetitive situations.

This talk is split into two parts. In the first part, features of the new NAO laser head are firstly introduced. With the help of this laser head, it is possible to accurately scan indoor environment and create the map from it. Secondly, some previous works about SLAM based on laser sensor will be presented.
From the past decades, many researchers attempt to solve the problem of simultaneous localization and mapping, commonly called SLAM.Many papers use different ways to implement or optimize SLAM methods, but most of them concentrate a small part of SLAM. With representative methods of each part of SLAM, the motivation of this talk is to help ourselves understand SLAM better and to explore future research possibilities arising from SLAM.

Saccades direct gaze direction to peripheral visual targets. Their magnitude is adapted throughout lifetime, and adaptation can be easily induced experimentally. An adaptation scheme that also works in newborns is unlikely to rely on sophisticated object recognition, memory, and geometrical difference computation. The lack of a geometrical model is also appealing to roboticists as it renders unnecessary any hardware-specific calibration/parameterization. We consider possible feedback signals that drive saccade adaptation, and differentiate two possible mechanisms. The first is a signed feedback, which contains information about whether the saccade was too long or too short. The second exploits the high foveal resolution, containing information about the resolution increase of the saccade target but not containing any vectorial error signal. We implement both mechanisms in a model and find experimental support in the first mechanism, at least in the case of horizontal saccades. Considering the separate neuroanatomical pathways for horizontal and vertical saccade control, we tentatively suggest the first mechanism to implement horizontal saccade adaptation and the second mechanism to implement vertical saccade adaptation.

Achieving cooperation among autonomous and rational agents is still a major challenge. In the past, altruistic cooperation was generally explained through genetic kinship relations. However, the theory of 'cultural kin' is an approach that tries to explain altruism through cultural relatedness. To promote cooperation among autonomous and rational agents, this work transfers the idea of cultural characteristics, which benefits social behaviour, to multiagent systems (MAS). Accordingly, agents are characterised by cultural traits, which they can imitate from their neighbours and are supposed to solve tasks, for which they need the cooperation of other agents in most cases. The interaction of cooperation and cultural trait propagation will be investigated in a theoretical analysis and in an empirical simulation in a particular developed framework. As a novelty, schemata will be analysed that are beyond the well-studied one-to-one interaction.

In this talk we are analysing the ensemble learning AdaBoost according to the example of the licence plate localization. We analyse the behaviour of the learning method using different classifiers which are combined to an ensemble. One disadvantage of AdaBoost as an ensemble-learning method is the long training duration. We can update AdaBoost a bit by making only minor changes. That way we are able to reduce a large quantity of information. Using less information as a basis we are able to train a good essemble classifier in a short time. To train the final ensemble classifier we use AdaBoost in its unmodified form. Finally we compare our results with two different cascades. One with an exhaustive search and one that was trained with a randomised limited attribute quantity. Using Threshold-, Binning-, FuzzyBinning classifiers and the "Joint Haar-like Feature" we demonstrate that our method is giving good results.

In this talk we study the potential of nature-inspired, in particular computational neural and symbolic representations, in order to build new adaptive knowledge-based systems, learning ambient intelligence systems, multimodal neural agents, self organizing knowledge discovery systems and human robot interaction systems. We will give an overview of learning neural symbolic knowledge technologies and robots from a perspective of integrative hybrid intelligent systems. We illustrate some new developments including also examples under development in the new Knowledge Technology lab.

The seminar is split into two parts. In the first half an extension of the classical backpropagation algorithm will be discussed and preliminary results will be shown. In the second part thoughts about a new project will be presented, hopefully culminating in a lively discussion with valuable comments and suggestions.
The classical backpropagation algorithm led to a "renaissance" in the artificial neuronal network community when it was (re-)introduced in the mid-1980s and has been used vastly until then. However, in recent years alternative methods have gained popularity and are often used to solve classification tasks or regression problems. This is due to different causes ranging from scalability issues to difficulties in interpreting the hidden layer. To possibly overcome some of these problems and enhance performance as well as the speed of convergence we suggest to extend the algorithm with the concept of intrinsic plasticity (IP), leading to a sparse firing pattern of the neurons. I will review classical backpropagation, elucidate the concept of IP and show preliminary results of the IP enhanced backpropagation algorithm.
In the second part of the talk I will present some thoughts for a new project. Using the concept of Sensorimotor Contingencies (SMCs) and its original idea, namely that "seeing is a way of acting", I want to discuss with you how to convey this notion to a NAO experiment giving rise to object discrimination.

Person tracking is an important topic in ambient living systems as well as in computer vision. Detecting a person from a ceiling-mounted camera is a challenge, since the person's appearance is very different from the top or from a side, and the shape of the person changes significantly when moving around the room. This paper proposes a novel algorithm for real-time person tracking using particle filters and different visual streams as cues for recognition. We present a new architecture to integrate different vision streams into a particle filter with adaptive reliability of each stream. In particular, a short-term memory mechanism is modeled, which enhances the robustness of the tracking system against noise.

Barsalou (2008) claims that conceptions of grounded cognition reject the standard view that amodal symbols represent knowledge in semantic memory. Furthermore it is unlikely that the brain contains amodal symbols; if it does, they work together with modal representations to create cognition. The talk in the Oberseminar will be a combination of: 1. A presentation of Lawrence W. Barsalou's paper 'Grounded Cognition' in a journal club fashion and 2. A discussion of implications for future research directions and the development of computational models of grounding language.
Reference: Lawrence W. Barsalou. Grounded Cognition. Annual review of psychology, 59:617-645, Annual Reviews, 2008

Multi-robot Exploration Behavior based on Mirror Neuron System by Junpei Zhong The mirror neuron system is a neuron system which is believed to play a role in action understanding and imitation among humans and other primate species. There have been a number of researchers who contributed to build up a computational functional model of mirror neuron system, mostly focused on motor action imitation. However, my research question will be at a higher level in action selection of multi-robot system and examines: how to select an action for robots coordination and how it supports the social organization between agents?
Humanoid robot behavior modulated by an artificial self-protective affective system by Nicolas Navarro Affective mechanisms are a result of a distributed neural system affecting several behavioral levels and for this reason there is no easy way to separate them from cognition. Moreover, affective mechanisms could be a requisite for intelligence, which is supported by neuroscientific insights of its key role in long-term memory and allocation of attention, two important characteristics ascribed to intelligent organisms. During years, studies have shown that the amygdala (an old brain structure) is involved in emotional evaluation of stimuli, and furthermore, it is involved in the associations of neutral stimuli with innate responses. Innate responses such as fear, are essential for the survival of any animal. For this reason, we believe they are an interesting start point to build an artificial intelligent system.