Memories of Artificial Neural Networks

Bjorn Schuller (TU Munich, Germany)

Abstract:

Artificial Neural Networks have recently undergone a certain renaissance given their manifold strengths and recent developments. They usually suffer from a lack of or vanishing memory - even when being recurrent. This comes, as the memory is typically not modelled explicitly. The long short term memory (LSTM) paradigm helps overcome this gap. Likewise, a network can actually "memorize", but it can also learn when to forget. This allows analysing the time span that has been learnt as being of merit.

In this vein, this talk deals with analysing "memories" of LSTM networks focussing on audio/visual processing tasks including best results in several recent research challenges. It further shows how to combine LSTM with dynamic Bayesian networks to reach dynamic abilities along-side explicit memory modelling.

Bio:

Björn Schuller received his diploma, doctoral degree, and habilitation in electrical engineering and information technology from TU Munich/Germany. He is a full professor heading the Institute for Sensor Systems at the University of Passau/Germany and the Machine Intelligence and Signal Processing (MISP) Group at TUM, and is the CEO of audEERING. Further, he is a permanent visiting professor at the Harbin Institute of Technology/P.R. China, and associate researcher of the University of Geneva/Switzerland and JOANNEUM RESEARCH in Graz/Austria. Previously, he was a guest lecturer at UNIVPM in Ancona/Italy, with the CNRS-LIMSI in Orsay/France, and visiting scientist of the Imperial College London/UK and NICTA in Sydney/Australia. Dr. Schuller (co-)authored 5 books and more than 300 publications (>4,000 citations, h-index = 33). He was co-founding steering committee member and guest editor, and still serves as associate editor of the IEEE Transactions on Affective Computing, the IEEE Transactions on Neural Networks and Learning Systems, the IEEE Transactions on Cybernetics, and Computer Speech and Language, and as guest editor for the IEEE Intelligent Systems Magazine, Neural Networks, Speech Communication, Image and Vision Computing, Cognitive Computation, and the EURASIP Journal on Advances in Signal Processing.