Self-organising neural network for binary image recognition

Most current work on neural networks involves supervised learning, where the provision of labelled training data are required. Since the target response of the network is known, the error between the target and current response of the network can be employed to adapt the strength of the network connections of synapses. The supervision of the learning process implies that the network is forced to comply with an already known model. This is, obviously, often not the case. To overcome these problems, it is necessary to use unsupervised learning, where the network is given no explicit labelling but learns to construct its own representations of the important primitive features in pattern space. The authors consider self-organisation which is a method of unsupervised learning in which neurons are adapted by a form of lateral inhibition. Similar pattern vectors are moved towards a target by a proportional change in the synaptic weights. This technique can generate topologically ordered feature maps. For successful recognition, similar patterns must be topologically close so that novel patterns are in the same general area of the feature map as the class they are most like