Title: Learning Requires Binding:

Overcoming the Context Deadlock

Christoph von der Malsburg

Computer Science Department and Program in Neuroscience

University of Southern California, Los Angeles and

Institut für Neuroinformatik, Ruhr-Universität Bochum

Abstract:

To understand the brain, four fundamental questions must be answered:

a) What is the data structure of brain states?

b) How are brain states organized?

c) What is the data structure of memory?

d) What is the mechanism of learning?

Underlying all these questions is the problem of how do brains self-organize. Or put another way, how do we learn from experience? Learning in artificial neural networks, for instance, is restricted to inputs of no more than about 100 bits per input pattern, whereas our senses deliver millions of bits per pattern. I will argue that the problem is due to a deadlock: signals are context-dependent, so that learning is not possible without prior recognition of context while recognition of context, seems to require prior learning. I will show how this deadlock can be broken by using a central binding mechanism and rapid reversible synaptic plasticity. I will conclude by discussing issues raised (at Stanford and elsewhere) that call temporal signal binding into question.