Modeling vision with the neocognitron

Abstract

The neocognitron, which was proposed by Fukushima [44.1], is a neural network model capable of robust visual pattern recognition. It acquires the ability to recognize patterns through learning. The neocognitron is a hierarchical network consisting of many layers of neuron-like cells. Its architecture was originally suggested from neurophysiological findings on visual systems of mammals. There are bottom-up connections between cells in adjoining layers. Some of these connections are variable and can be modified by learning. The neocognitron can acquire the ability to recognize patterns by learning. Since it has a large power of generalization, presentation of only a few typical examples of deformed patterns (or features) is enough for learning. It is not necessary to present all of the deformed versions of the patterns that might appear in the future. After learning, the neocognitron can recognize input patterns robustly, with little effect from deformation, changes in size, or shifts in location. It is even able to correctly recognize a pattern that has not been presented before, provided that it resembles one of the training patterns. The principle of the neocognitron can be used in various kinds of pattern recognition systems, such as recognizing handwritten characters. Further extensions and modifications of the neocognitron have been proposed to endow it with a function of selective attention, an ability to recognize partly occluded patterns, and so on.