Deep architectures and folding¶
Lech Szymanski
Doctor
University of Otago
Abstract: There is no doubt that in recent years deep neural networks have pushed
the envelope of the state of the art machine learning. But, as is typical of
artificial neural network models, a huge share of this success is due to the
uncanny, often unexplainable, intuition of the expert users, who make
decisions about the network architecture for a given problem. We need to
understand these models at a far deeper level if we're going to push the
performance envelope even further. In this talk I will present my research
on the fundamental differences between internal representation in shallow
and deep architectures, with the aim of establishing how and when the
latter can be better. Specifically, I will discuss one type of deep
representation that is analogous to the folding of the input space, and how
effective it can be at approximation of functions with repeating patterns.
Bio: Dr Lech Szymanski is a lecturer at the department of Computer Science,
University of Otago, New Zealand. His main research interests include
machine learning, deep representations and connectionist models,
especially with applications to computer vision. Before his PhD, which he
completed in 2012, he worked as a software engineer for a wireless
telecommunications company in Ottawa, Canada.