Jump to main content

Professorship Artificial Intelligence

Research Seminar

The seminar is intended for interested students in their master or final stages of bachelor studies. However, other interested listeners are cordially welcomed too! In the seminar, either the staff of the professorship Artificial Intelligence or students will present current research topics. The presentations are normally hold in English. The seminar will be hold irregularly in the room 1/336 whereby the precise schedule is presented in the following table.

Information regarding diploma and master students

Seminar presentations of each major could also be hold in the frame of this event. This includes mainly the seminar “Hauptseminar" in the major Diplom-IF/AIF or the "Forschungsseminar" in all master majors of Informatik. Both courses have the goal that the attendees acquire research-oriented knowledge by themself and present it during talks. The research topics will be typically from the field of Artificial Intelligence with emphasis on object recognition, neuro-computing at graphic cards or multi-core-CPUs, reinforcement learning and intelligent agents in virtual reality. However, other topics could also be chosen. Interested students should write an Email to Prof. Hamker, the talk itself will be scheduled after an individual consultation.

Next Events

Investigating robustness of End-to-End Learning for self-driving cars using Adversarial attacks

Vivek Bakul Maru

Mon, 27. 5. 2019, 14:00, Room 131

Deep neural networks have been recently performing on high accuracy in many important tasks, most notably on image-based tasks. Current state-of-the-art image classifiers are performing at human level. Having said that, Deep learning has been successfully applied to End-to-End learning architectures for autonomous driving tasks where a deep neural network maps the pixel values coming from camera to a specific steering angle of the car. However, these networks are not robust enough to carefully perturbated image inputs known as Adversarial inputs. These inputs can severally decrease the accuracy and performance of the network which results in endangering the systems where such deep learning models are deployed. This thesis work mainly focuses on proving the vulnerability of End-to-End architectures for self-driving car towards adversarial attacks. In this work, we formulize the space of adversaries against End-to-End learning and introduce a way to generate Universal Adversarial Perturbation for these architectures. We have generated universal perturbations using distinct subsets of dataset and tested against two different models i.e. classification and regression in a White-box setting. To show the transferability and universality of these perturbations, we have tested these perturbations against a third constrained model in a Black-box setting. With the results, we showed that End-to-End architectures are highly vulnerable to perturbations and these perturbations can generate a serious threat to the system when deployed in a self-driving car. We also propose a theoretical prevention method for the End-to-End architecture against Universal perturbations to increase the security of the model.


Hoa Tran Duy

Mon, 3. 6. 2019, 14:00, Room 131


Discomfort Detection in Automated Vehicles Using Gaze data and Machine Learning Methods

Roghayyeh Assarzadeh

Wed, 12. 6. 2019, 14:00, Room 132


Recent Events

Successor representations

Julien Vitay

Mon, 20. 5. 2019, Room 131

Successor representations (SR) are a trade-off between model-free and model-based methods in reinforcement learning. They see a revival for a couple of years through the work of Samuel Gershman and colleagues who propose a neurobiological mapping of this algorithm to the prefrontal cortex, hippocampus and basal ganglia. This seminar will explain the principle of SR, present their main neuroscientific predictions and discuss their plausibility. See

Development and analysis of active SLAM algorithms for mobile robots

Harikrishnan Vijayakumar

Mon, 13. 5. 2019, Room 375

Mobile robots are becoming an essential part of industrial and household services. In this thesis, we tackle the problem of autonomous exploration using active SLAM concepts. Autonomous exploration is the task of creating a model of the environment by the robot without human intervention. The environment model creation capabilities of the robot are highly dependent on sensor and actuator performance. We require an algorithmic approach to handle the noise from the hardware measurements. We need a combined solution for localization, mapping, and motion planning to improve autonomous exploration performance. In its classical formulation, SLAM is passive which means that the algorithm creates the map and localizes the robot passively by using received motion and sensor observations. It takes no part in deciding where to take measurements next. The quality of the map created by the SLAM depends on the correctness of position estimation of the robot while mapping. To further improve the quality of the map, it is necessary to integrate path planning with SLAM so that the algorithm can decide the future motion of the robot based on the current localization and map information for achieving better localization and a higher quality map. This thesis addresses this problem. There are many approaches that do not address the uncertainty in motion and sensor measurements. We propose a probabilistic approach that considers the current system uncertainty and plans the path of the robot that minimizes the robot?s pose and map uncertainty. This thesis first examines the background and other approaches addressing active SLAM. We implement and evaluate state of the art approaches on our own, standardized, datasets. We propose and implement a new approach based on Schirmer on path planning in belief space. Path planning in configuration space assumes accurate knowledge of the robot position, and the goal is to find a path from the position of the robot to the destination. In reality, uncertainties in movement and sensing sometimes lead to the inability to detect the accurate position of the robot. Planning in belief space maintains a probability distribution over possible states of the robot and takes these uncertainties into account. Path planning in belief space entails planning a path for the robot given the knowledge about the environment and the current belief. We compute paths in belief space by computing two representations of the environment in parallel. The first one is a classical geometric map of the environment, the other one encodes how well the robot can localize at different parts of the environment. We then apply a search algorithm to the discrete configuration space to find the best path to the target goal location with the lowest uncertainty. Finally, we define a utility function for each target goal based on its mean uncertainty and navigation cost to evaluate the best candidate for the next motion. At the end of the work, we discuss in detail the results, strengths, weaknesses, assumptions, and identify areas for future research.

Optimizing topic models on small corpora - A qualitative comparison between LDA and LDA2Vec

Anton Schädlich

Thu, 9. 5. 2019, Room 367

Topic modeling become a subfield in unsupervised learning since the invention Latent Dirichlet Allocation (LDA) algorithm. However, on smaller corpora it sometimes does not perform well in generating sufficiently coherent topics. In order to boost coherence scores, it has been extended in various researches with vector space word embeddings. The LDA2Vec algorithm is one of these symbiotic algorithms that draws context out of the word vectors and the training corpus. This presentation is about the qualitative comparison of the topics and models of optimized LDA and the LDA2Vec algorithm trained on a small corpus of 1800 German language documents with a considerably small amount of topics. The coherences were measured both on the 'combined coherence framework' and on manual observation.

Multi-GPU simulation of spiking neural networks on the neural simulator ANNarchy

Joseph Gussev

Mon, 6. 5. 2019, Room 367

Because of their highly parallel architecture and availability, GPUs have become very present in neural simulators. With the goal of outperforming CPU-based clusters with cheaper GPUs, some simulators have started to add multi-GPU support and have shown speedups compared to single-GPU implementations. The thesis behind this presentation found its goal in adding multi-GPU implementations to ANNarchy's spiking neural network simulation. After a brief background overview, the ideas behind the implemented algorithms will be presented, their performance and scalability evaluated and discussed.

A model car localization using Adaptive Monte Carlo Localization and Extended Kalman Filter

Prafull Mohite

Tue, 23. 4. 2019, Room 336

Localization works with the help of GPS. But this might not be very accurate specially in case of autonomous driving. In case of Autonomous driving, It is a key step for deciding what should be next step. As localization is a key concept to predict the next step, it should be made more accurate and reliable and this can be done using Exteroception sensors. The range of Lidar is high enough with minimum error rate, is been used for depth perception. In this thesis, a very popular technique, Simultaneous Localization and mapping (SLAM) approach is selected. Map is needed for localization and localization cannot be completed without map giving rise to chicken egg problem. Mapping the environment before localization is in common practice and we have followed that. The aim of thesis is to tryout the performance of the proposed system against the state of the art. The proposed system consist of hybrid localization consist of 3 times pose correction with non Gaussian filter and Nonparametric filters. As a result, location in Cartesian coordinate system and orientation is observed.

Investigation of Model-Based Augmentation of Model-Free Reinforcement Learning Algorithms

Oliver Lange

Mon, 15. 4. 2019, Room 131

Reinforcement learning has been successfully applied to a range of challenging problems and has recently been extended to handle large neuronal network policies and value functions. However, the sample complexity of model-free algorithms tends to limit their applicability to physical systems. The reduction of sample complexity for deep reinforcement learning algorithms is essential to make them applicable in complex environments. This presentation is about the investigation of the performance of a reinforcement learning agent which utilizes the advantages of both model-free and model-based reinforcement learning algorithms.

Interpreting deep neural network-based models for automotive diagnostics

Ria Armitha

Wed, 10. 4. 2019, Room 336

With the breakthrough of Artificial intelligence over the last few decades and extensive improvements in Deep Learning methodologies, the field of Deep Learning has gone through major changes. AI has outdone humans in computing complex tasks like object and image recognition, fault detection in vehicles, speech recognition, medical diagnosis etc. From a bird's-eye view the models are basically algorithms which try to learn concealed patterns and relationships from the data fed into it without any fixed rules or instructions. Although these models' prediction accuracies may be impressive, the system as a whole is a black-box (non-transparent). Hence, explaining the working of a model to the real world poses its own set of challenges. This work deals with interpreting vehicle fault-detection model. Current fault detection approaches rely on model-based or rule-based systems. With an increasing complexity of vehicles and their sub systems, these approaches will reach their limits in detecting fault root causes in highly connected and complex systems. Furthermore, current vehicles produce rich amounts of data valuable for fault detection which cannot be considered by current approaches. Deep Neural Networks (DNN) offer great capabilities to tackle these challenges and automatically train fault detection models using in-vehicle data. However, fault detection models based on DNNs (here, CNNs and LSTMs) are black boxes so it is nearly impossible to back-trace their outputs. Therefore, the aim of this work is to identify, implement and evaluate available approaches to interpret decisions made by DNNs applied in vehicle diagnostics. With that, decisions made by the DNN diagnostics model can be better understood to (i) comprehend the model's outputs and thus increase model performance as well as (ii) enhance their acceptability in vehicle development domain.

Extrinsic Camera Pose Estimation Using Existing Parallel Lines of the Surrounding Area

Layth Hadeed

Mon, 8. 4. 2019, Room 131

Camera pose estimation is an essential process for computer vision systems used by intelligent automotive systems. This work is concerned with the pose estimation of a train camera that is installed in the front windshield of the train. The pose is estimated by using only the rails and sleepers of the railway track without any additional calibration objects. The information extracted from the rail lines in the image helped estimate the pitch and yaw by determining the vanishing point in the world X-axis direction. The orthogonality between the rails and sleepers helped estimate the roll by using the orientation of the sleepers in the track, it was possible to estimate the value of the role. The translation parameters are calculated from the rotation matrix and the normal vectors of the projective planes between the camera center and the rails. The results show that a good approximation of the pose can be made by adding an additional step that compensates for the error in the estimated roll.

Weight estimation using sensory soft pneumatic gripper

Hardik Sagani

Mon, 25. 3. 2019, Room 132

Current soft pneumatic grippers might robustly grasp flat and flexible objects with curved surfaces without distorting them, but not estimate the properties by manipulating it to a certain extent. On the other hand, it is difficult to actively deform a complex gripper to pick up surface objects. Hence, this research project represents a prototype design of soft sensory pneumatic gripper utilized to estimate the weight of objects. An easy to implement soft gripper is proposed by soft silicon material with four fingers. It can be fabricated by placing piezoresistive material sandwiched between soft silicon rubber and conducting threads. Sixteen Pressure sensors and 4 curvature sensors of velostat piezoresistive material have been placed into the hand. The layers were designed with the help of conductive thread electrode material. Hence, the sensory gripper can evaluate the weight property of grasped objects.

Neuro-computationale Modellierung der Basal Ganglien für Stopp-Signal Aufgaben

Iliana Koulouri

Tue, 12. 3. 2019, Room 132

Baladron, Nambu und Hamker haben 2017 ein biologisch motiviertes neuro-computationales Basalganglienmodell für die Aktionsauswahl während einer Erkundungsphase nach Änderung von Umweltkonditionen durch die GPe-STN Schleife entwickelt. Die Basalganglien beeinflussen ebenso die Unterdrückung von Handlungen. Es wird angenommen, dass ein Rennen zwischen Go-Signalen und Stopp-Signalen einem Rennen zwischen den neuronalen Bahnen der BG entspricht. Dazu postulierten Schmidt et al. (2013) und Mallet et al. (2016) ein zweistufiges Abbruchmodell, welches einen Pause- und einen Abbruchmechanismus beinhaltet. Der Abbruch einer Handlung werde durch einen neu entdeckten Zelltyp, den arkypallidalen globus pallidus Zellen, realisiert. Über ihre Vernetzung sind bislang nicht alle Einzelheiten bekannt. Die vorliegende Arbeit befasst sich mit der Erweiterung des BG Modells von Baladron et al. (2017) um die Implementierung dieses postulierten zweistufigen Abbruchmodells, der Replizierung der Ergebnisse von Mallet et al. (2016) und der Untersuchung des Einflusses der Stärke der Inputsignale, sowie ihre Verzögerung auf das Modell. Es zeigt sich, dass das erweiterte Modell die Ergebnisse von Mallet al al. (2016) erfolgreich replizieren kann. Dabei scheint auch der Kortex als Quelle der Erregung der arkypallidalen globus pallidus Zellen als plausibel. Das Modell weist einen geringen Einfluss gegenüber Änderungen der Stoppraten und der zeitlichen Verzögerung zwischen Go- und Stopp-signalen auf. Allerdings erweist es sich als empfindlich gegenüber Änderungen der Goraten.

Modelling and controlling of the soft pneumatic actuator for the Elbow assistance

Sidhdharthkumar Vaghani

Tue, 12. 3. 2019, Room 132

More than a decade the wearable robotics is in the trend for the rehabilitation or for the assistance of human motion. The wearable robots are attracting the focus of the human because of its safe interaction with the human and external environment. Because of their stiffness and wearable property, wearable robots are very famous for the rehabilitation and assistance of human body joints e.g. knee rehabilitation, elbow or shoulder assistance etc. However the controlling of the soft robots is not an easy task compared to rigid robots. Mainly the wearable robots are using the air pressure as the activation parameter of their body. Therefore the aim of this thesis is to modelling and controlling of the exosuit, which is useful for the elbow assistance. The structure of the exosuit consists of two pneumatic actuator one is for elbow flexion movement and other is for extension movements. One IMU unit is mounted on the top of the exosuit to monitor the relative angle between two links of the exosuit. The modelling of the exosuit is carried out separately for flexion and extension actuator by considering the geometry of the actuators. For precise control of the actuator proportional valve has been developed and used along with the PID controller. The multi-layer multi-pattern central pattern generator (MLMP-CPG) is adapted to generate the rhythmic and non-rhythmic behaviour of the exosuit. To implement a better control strategy, the reflexes are introduced with the CPG model. At the end of the thesis work, the experiments are carried to prove the implementation of PID and reflex based controller using the MLMP-CPG. The reflexes with the CPG generate the adaptive behaviour of the exosuit.

Decorrelation and sparsification by homeostatic inhibitory plasticity in cooperation with voltage-based triplet STDP for receptive field formation

René Larisch

Mon, 4. 3. 2019, Room 202

Several recent models have shown how V1 simple-cell receptive fields may emerge from spike-timing dependent plasticity (STDP) mechanisms. These approaches mostly depend on normative learning paradigms, such as maximizing information transmission or optimizing sparseness, or on simplified rate-based learning dynamics. These assumptions make it unclear how inhibition influences neuronal properties (for example orientation selectivity and contrast invariance tuning curves) and network effects (for example input encoding). We here propose a model of V1 based on a phenomenological excitatory and an inhibitory plasticity rule. Our simulation results showing the interplay of excitatory and inhibitory plasticity on the connectivity structure and the influence on input encoding. Further, we demonstrate how the right amount of excitatory and inhibitory input leads to the occurrence of contrast invariance tuning curves.

A network model of the function and dynamics of hippocampal place-cell sequences in goal-directed behavior

Lorenz Gönner

Fri, 15. 2. 2019, Room 336

Hippocampal place-cell sequences observed during awake immobility often represent previous experience, suggesting a role in memory processes. However, recent reports of goals being overrepresented in sequential activity suggest a role in short-term planning, although a detailed understanding of the origins of hippocampal sequential activity and of its functional role is still lacking. In particular, it is unknown which mechanism could support efficient planning by generating place-cell sequences biased toward known goal locations, in an adaptive and constructive fashion. To address these questions, I propose a spiking network model of spatial learning and sequence generation as interdependent processes. Simulations show that this model explains the generation of never-experienced sequence trajectories in familiar environments and highlights their utility in flexible route planning. In addition, I report the results of a detailed comparison between simulated spike trains and experimental data, at the level of network dynamics. These results demonstrate how sequential spatial representations are shaped by the interaction between local oscillatory dynamics and external inputs.

Computing with hippocampal sequences

Prof. Dr. Christian Leibold

Fri, 15. 2. 2019, Room 336

Hippocampal place cells are activated in sequence on multiple time scales during active behavior as well as resting states. Reports of prewired hippocampal place cell sequences that were decoded as future behaviors are hard to bring into register with the suspected roles of the hippocampus in the formation of autobiographic episodic memories and spatial learning. Here, I propose a computational model, that shows how a set of predefined internal sequences that are linked to only a small number of salient landmarks in a large random maze can be used to construct a spatial map of a previously unknown maze.

Adaption of a neuro-computational model of space perception to new physiological data and evaluation with experimental data from human and monkeys

Nikolai Stocks

Wed, 13. 2. 2019, Room 132

A. Zietsche and F.H. Hamker developed a computational model for the simulation visual perception and perisaccadic mislocalisation [A. Ziesche et al. 2011]. .A paper by Joiner et al. published in 2013 demonstrates that monkeys do perform different than humans in Saccadic Supression of Displacement (SSD) trials under the stimulus blanking condition which the model does not account for. Furthermore does data by Xu et al. 2012 show that the neural representation of current eye position updates later than the original model accounted for. The goal of this thesis is to find adjustments to the parameters of the original model and to allow it to accurately simulate SSD experiments for both monkeys and humans using Xu's revised timeframe.

Development of Collision Prevention Module using Reinforcement Learning

Arshad Pasha

Wed, 6. 2. 2019, Room 132

The tremendous advancements in the field of application of Artificial Intelligence in Advanced Driver Assistance Systems (ADAS) and Autonomous Vehicles (AVs) have still not been able to curb the accident rate. Many automotive manufacturers aim to achieve the 'Zero Fatalities' goal with advanced collision prevention modules by the next decade. This thesis work mainly focuses on the implementation of ADAS functionalities such as Autonomous Emergency Braking (AEB), Autonomous Emergency Steering (AES), and Adaptive Cruise Control (ACC) using Reinforcement Learning (RL). RL has attracted researchers' attention in recent years due to its immense potential. Q-Learning is one of the most exploited RL algorithms. It has not only been employed to solve a simple problem as grid world problem but also to solve complex problems as playing advanced levels video games such as Doom, Breakout and various other Atari games. This thesis work also uses the Q-learning algorithm. This thesis was carried out at the Research and Development unit of ZF known as Zukunft Mobility GmbH, Kösching, Germany. The principal objective of this thesis is to use Tabular Q-learning approach to implement complex ADAS functionalities as mentioned earlier and analyze the RL agent's behavior in addition to integrating these developed algorithms within the organization's proprietary simulation environment (SimpleSpec). Further, the continuous State-Action space has been discretized to adapt to the tabular Q learning approach which avoids the 'Curse Of Dimensionality' and 'Matrix Explosion' problems. In the penultimate part, a qualitative and quantitative evaluation of the performance has been carried out. The results obtained after a full-scale model testing for low, medium and high-speed scenarios have been recorded. In addition, random scenarios were generated to test the scope and capabilities of the trained RL agent. Further research has been carried out in the area of the impact of reward function structure on the RL agent which includes the sparse and non-sparse reward structure design and their influence on the reward convergence and learning process. Further, this segment also includes the analysis between the explicitly distributed non-sparse reward design and the non-sparse inline reward design. The final part summarizes the tasks covered, the goal achieved and draws a path to the future work.

Road scene semantic segmentation using residual factorized convnet and surround view fisheye cameras

Saiffuddin Syed

Wed, 30. 1. 2019, Room 367a

Automotive industry is continuously evolving, especially in the self-driving domain which creates a demand for new concepts to be developed, implemented and tested. At present the only sensor capable of sensing the immediate surrounding of the vehicle is a camera.This thesis addresses the 360 degrees road scene semantic segmentation problem for fisheye cameras. Present vehicles are equipped with distinct types of cameras used for various practical real-time applications, the most common camera model being the wide-angle fisheye cameras which are considered for this thesis. Usage of this camera brings two major challenges: firstly, CNN-based semantic segmentation task requires a huge amount of pixel-level annotated data. So far there is no open-source annotated dataset available for wide-angle images. Secondly, a fisheye camera introduces severe distortions and negates the positional invariance offered by a conventional pinhole camera model. To overcome this, training the images on transformed images that are augmented using a fisheye filter is proposed. An approach to integrate blocks which improve the representational power of existing architectures by explicitly modelling interdependencies between channels of convolutional features, has been tested. The experiments carried out prove the effectiveness of these blocks when augmented data is used. Much of the work presented in the thesis was devoted to a rigorous comparison of the architectures.The evaluation of the thesis is done on two different kind of datasets, a real world dataset and a synthetic dataset. The primary metric used for evaluation was the Intersection-over-Union (IoU). The results at the end of the thesis showed that a large amount of existing annotated data taken from pinhole cameras can be reused through augmentation and relatively small amount of annotated from fisheye cameras is required to account for domain shift. Further, the new architectures presented in this thesis show promising results when applied to augmented data.

Categorizing facial emotion expressions with attention-driven convolutional neural networks

Valentin Forch

Mon, 21. 1. 2019, Room 219

The development of so-called deep machine learning techniques has brought new possibilities for the automatic processing of emotion-related information which can have great benefits for human-computer interaction. Vice versa machine learning can profit from concepts known from human information processing (e.g., visual attention). Being located in the spectrum of human and artificial intelligence, the aim of the present thesis was twofold: (a) to employ a classification algorithm for facial expressions of emotions in the form of a deep neural network incorporating a spatial attention mechanism on image data of facial emotion expressions and (b) to compare the output of the algorithm with results from human facial emotion recognition experiments. The results of this thesis show that such an algorithm can achieve state-of-the-art performance in a facial emotion recognition task. With regard to its visual search strategy some similarities with human saccading behavior emerged when the model's perceptive capabilities were restricted. However, there was only limited evidence for emotion-specific search strategies as can be found in humans.

A network model of the function and dynamics of hippocampal place-cell sequences in goal-directed behavior

Lorenz Gönner

Mon, 21. 1. 2019, Room 219

Hippocampal place-cell sequences observed during awake immobility often represent previous experience, suggesting a role in memory processes. However, recent reports of goals being overrepresented in sequential activity suggest a role in short-term planning, although a detailed understanding of the origins of hippocampal sequential activity and of its functional role is still lacking. In particular, it is unknown which mechanism could support efficient planning by generating place-cell sequences biased toward known goal locations, in an adaptive and constructive fashion. To address these questions, I propose a spiking network model of spatial learning and sequence generation as interdependent processes. Simulations show that this model explains the generation of never-experienced sequence trajectories in familiar environments and highlights their utility in flexible route planning. In addition, I report the results of a detailed comparison between simulated spike trains and experimental data, at the level of network dynamics. These results demonstrate how sequential spatial representations are shaped by the interaction between local oscillatory dynamics and external inputs.

3D reconstruction with consumer depth cameras

Manh Ha Hoang

Wed, 9. 1. 2019, Room 132

In this thesis, we develop an RGB-D camera-based system that is able to generate a 3D model of a single household object using a consumer depth (RGB-D) camera. The system then grabs textures of the object from a high-resolution DSLR camera and applies them to the reconstructed 3D model. Our approach specially addresses on generating a highly accurate 3D shape and recovering high-quality appearance of the object within a short time interval. The high-quality 3D texture object models can be used for the products of online shopping, augmented reality, and further research of 3D Machine Learning.

Hierarchical representations of actions in multiple basal ganglia loops

Javier Baladron

Wed, 5. 12. 2018, Room 132

I will introduce here three novel concepts, tested and evaluated by means of a neuro-computational model that brings together ideas regarding the hierarchical organization of the basal ganglia and particularly assigns a prominent role to plasticity. I will show how this model reproduces the results of two cognitive tasks used to measure the development of habitual behavior and introduce a model prediction.

Investigating reservoir-based reinforcement learning for robotic control

Oleksandr Nikulkov

Wed, 28. 11. 2018, Room 132

Reservoir Computing is a relatively novel approach for training recurrent neural networks. It is based on generating a random recurrent reservoir as a part of the network and training only the readout of the reservoir. This separation makes the setup easy to be implemented and offers different directions for further research to be done. Existing methods for learning cognitive tasks often require continuous reward signals, which are not always available in cognitive tasks. However, this disadvantage can be avoided by using supralinear amplification on the trace of node-perturbation weight updates to suppress the relaxation-effect, as proposed by (Miconi, 2017). In this presentation, I will show how such a network can be applied to a robotic control task and investigate the role of the different parameters.

... older

Press Articles