A new paradigm for the classification of patterns: The 'race to the attractor' neural network model.

Abstract

The human brain is arguably the best known classifier around. It can learn complex classification tasks with little apparent effort. It can learn to classify new patterns without forgetting old ones. It can learn a seemingly unlimited number of pattern classes. And it displays amazing resilience through its ability to persevere with reliable classifications despite damage to itself (e.g., dying neurons). These advantages have motivated researchers from many fields in the quest to understand the brain in order to duplicate its ability in an artificial system. And yet, little is known about the way the brain really works. But one fact which is apparent from available data is that 'TIME' is a critical component of its computational process. The brain is a dynamical system whose state evolves with time. Outside stimulus is processed and transformed repeatedly within it, with a multitude of signals interacting with each other in a complex, time-dependent manner. As a result, the process of pattern recognition inside the brain is also a time-dependent evolution of states where the initial image of the unknown pattern is progressively transformed into a form which represents the class of that pattern. In this thesis, we seek to achieve some of the advantages of the brain as a classifier by defining a model which captures the importance of time in the recognition process. The 'race to the attractor' neural network model involves the use of dynamical systems which transform initially unknown patterns into simpler prototypes which each represent a pattern class. The time required for this transformation to occur increases as the resemblance between the unknown pattern and the class prototype decreases. This results in a race where dynamical systems compete to transform the unknown pattern as quickly as possible. The winner of this race identifies the unknown pattern as a member of the class which the prototype of that dynamical system represents.