SLIDES HERE:
[ Ссылка ]
PhD Thesis Defense. Cornell, April 2016.
Title: Training and Understanding Deep Neural Networks for Robotics, Design, and Perception.
Abstract:
Artificial Neural Networks (neural nets) are a powerful class of models with both theoretical and practical advantages for many problems of interest. Networks with more than one hidden layer ("deep" neural nets) compute multiple functions on later layers that share the use of intermediate results computed on earlier layers. This compositional, hierarchical structure entails a strong model bias that works curiously well on real-world problems.
In this talk I will summarize several contributions to neural net research. First, I will show examples of how networks may be trained on two real-world problems for which the cost function is not differentiable: generating fast gaits for walking robots and representing and designing 3D shapes using crowdsourced human preferences. Second, I will discuss several studies that shed light on the inner workings of image classification networks trained using more traditional differential costs. These latter studies reveal some surprising features of large networks and lead to a greater understanding of and intuition for the computations performed by neural nets.
Thesis committee: Hod Lipson (chair), Shimon Edelman, Gun Sirer, Jeff Clune, Yoshua Bengio
