At the end of the module students have extensive theoretical knowledge of advanced deep learning architectures and their applications in robotics. Special attention is put on deep reinforcement learning, learning from small sample sizes and robustness evaluation.
Description
The lectures will provide extensive theoretical aspects of neural networks and in particular deep learning architectures, specifically for advanced methods in the field of Robotics, esp. deep reinforcement learning.
- Recap of deep learning in neural networks (multilayer perceptrons, CNN, automatic differentiation, optimization and regularization methods)
- Self-supervised learning
- Deep reinforcement learning (Bellman equation, Deep Q-Learning Deep Deterministic Policy Gradients, Trust Region Policy Optimization)
- Advanced deep reinforcement learning (attention mechanisms, Neural Turing Machines, Alpha Go, Alpha Zero, ...)
- Transfer and One Shot Learning (Siamese Networks, Progressive Neural Networks, combining simulated and real world samples)
- Network architectures guaranteeing robustness and providing confidence values for predictions; analysis of learned models
- Robotic applications (learning to grasp; tactile material classification; fast motion planning)
- Software frameworks for advanced deep learning (TensorFlow, Keras, Deepmind Sonnett, Facebook Torch)
- Open problems in Deep Learning for Robotics
Prerequisites
MA0902 Analysis für Informatiker MA0901 Lineare Algebra für Informatiker
IN2349 Deep Learning in Robotics
or
IN2346 Deep Learning for Computer Vision
Teaching and learning methods
The contents of the module is presented as a lecture with slides. The understanding is enhanced by case studies of learning problems from current robotics research.
Examination
The written exam (90min) at the end of the course makes sure the students have an understanding of the mathematical foundations of Deep Learning with Neural Networks, and of the special requirements in the application to robotic systems, such as reinforcement learning, evaluation of the robustness and small sample sizes.
Recommended literature
I. Goodfellow, Y. Bengio and A. Courville. Deep Learning. MIT Press, 2016. (http://www.deeplearningbook.org) Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. Kevin Murphy. “Machine Learning: A Probabilistic Perspective”, MIT Press 2012
The lectures will provide extensive theoretical aspects of neural networks and in particular deep learning architectures, specifically for advanced methods in the field of Robotics, esp. deep reinforcement learning.
• Introduction/Motivation
• Introduction to Machine Learning: A Bayesian View
• Advanced Network Architectures
• Generalization & Robustness: Adversarial Training
• Bayesian Deep Learning
• Transfer & Semi-Supervised Learning
• Learning Generative Models: GANs & Autoencoder
• Reinforcement Learning I
• Reinforcement Learning II
• Reinforcement Learning III
• Selected Robotic Applications
Material
• I. Goodfellow, Y. Bengio and A. Courville. Deep Learning. MIT Press, 2016.
• Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
• Kevin Murphy. “Machine Learning: A Probabilistic Perspective”, MIT Press 2012
