|  Lecturer  | 
			 Berthold Bäuml  | 
		
		
			|  Module  | 
			 IN2349  | 
		
		
			|  Type  | 
			 Lecture  | 
		
		
			|  Semester  | 
			 SS 2017  | 
		
		
			|  ECTS  | 
			 2.0  | 
		
		
			|  SWS  | 
			 2V  | 
		
		
			|  Audience  | 
			 Elective course in Robotics, Cognition, Intelligence (Master's Program)  | 
		
		
			|  Time & Place  | 
			 Do 12 - 14 MI HS 2  | 
		
	
 News 
 Description 
The lecture covers the mathematical foundations and the efficient implementation of modern Deep Learning Neural Network Architectures (incl. One-Shot Learning) and its application to problems with real robots (e.g., tactile material classification with a robotic hand).
 
-  Introduction/Motivation slides
  -  Machine Learning Basics I: slides 
-  linear algebra, probability & information theory, continuous optimization
 
 
  -  Machine Learning Basics II:   slides 
-  Revisiting Bayesian probabilistic inference, 
bias-variance decompostion, 
classifictation & logistic regression
 
 
  -  (Deep) Neural Networks I: slides 
-  XOR problem; universal function approximator; backpropagation; TensorFlow
 
 
  -  Deep Neural Networks II:  slides 
-  mutlilayer MNIST, TensorFlow/TensorBoard, activation functions, weight initialization, regularization
 
 
  -  Deep Neural Networks III:  slides 
-  adversarial training; convolutional neural networks, examples: AlexNet, ...
 
 
  -  Deep Learning IV and Application in Robotics:  slides 
-  CNN applications; transfer learning; gradient descent revisited; momentum methods
 
 
  -  Deep Learning V and Application in Robotics:   slides 
-  cross-validation; hyperparameter search; applications in robotics: grasping
 
 
 
 
 Course Materials 
 Literature 
  
 
-  Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
  -  Kevin Murphy. “Machine Learning: A Probabilistic Perspective”, MIT Press 2012
 
 
 Example Programs 
  
 Software Tools