After having conquered the world of pattern recognition, neural
networks have turned their attention to signal processing problems,
where they are seen to provide results that are comparable to or
better than those obtained with conventional signal-processing
solutions. In this talk we will attempt to understand the reason for
their effectiveness. We explore the simple principles that networks
build on, including their interpretations as boolean networks,
correlation filters, linear and nonlinear classifiers capable of
modelling non-convex classes, universal function approximators and
dictionaries of overcomplete repesentations.  We will discuss the
effect of these interpretations on network optimization and training.
Finally, we will look at a variety of neural network approachs for
signal enhancement, signal separation, and multi-channel signal
processing, and relate these to our interpretations.

Bhiksha Raj is an associate Professor at the Language Technologies Institute in the School of Computer Science at Carnegie Mellon University (CMU), Pittsburgh, USA, with additional affiliations to the Electrical and Computer Engineering and Machine Learning departments. Dr. Raj obtained a PhD from Carnegie Mellon in 2000. Between 2001 and 2008 he was a research scientist at Mitsubishi Electric Research Labs in Cambridge, MA, and has been at CMU since 2009. Dr Raj's research interests lie at the conjunction of speech and audio processing, natural language processing, privacy, and machine learning, with a more recent focus on derivation of information from large amounts of data.