Pattern Recognition Versus Synchronization: Pattern Recognition

Abstract

In previous chapters the synchronization between axonal pulses in pulsecoupled neural networks was in the foreground of our study. Pulse synchronization is, indeed, a well-established experimental fact. Occasionally, in Sects. 6.4 and 6.5 we had a look at the capability of such a network to recognize patterns. The network we studied could perform pattern recognition only to some extent, especially it was not able to fully distinguish between patterns. On the other hand, in the past, several computer models of neural networks have been developed aiming at more detailed pattern recognition. The corresponding networks do not use spiking model neurons, but rather concentrate on the concept of attractors. In such a case, the total state of the system at each moment is represented by a state vector in a high-dimensional space that refers to the activities of all neurons. The temporal evolution of the state vector is described by a set of differential equations which describe the time evolution of neuronal activity due to the internal dynamics (which are assumed to be very simple) and their interactions. The dynamics or, in other words, the temporal change of the state vector, can finish in stable fixed points which are a special kind of attractor. If in the state space the state vector comes close to a fixed point, it is attracted into that point, which means that a specific pattern is recognized.