UNSUPERVISED MODELS

Associative memory networks are designed for the best recovering the original noise-free pattern from an incomplete or distorted signal. The main characteristic of thefixed-weight association networks, is that the synaptic weights are precomputed and prestored. We shall review the Feedforward Associative Memory Networksand Feedback Associative Memory Networks, namely the Hopfield models.

The fixed-weight models have limited applications since they cannot adapt to changing environments. There is another variety of unsupervised networks, called Competitive Learning Networks , whose synaptic weights adapt according to unsupervised learning rules. These models can learn in the absence of teacher's guidance. In other words, the training is based exclusively on the input training patterns. The class of competitve leaning networks includes, for example, self-organization network, adaptive resonance, and Neocognitron. Self-organization and adaptive resonance approaches incorporate lateral inhibitions into the cooperative-competitive neural models.

Contents