Implementation of a 3-D model for neocognitron

Neocognitron is a widely used model to emulate the human visual system. In this paper, the implementation of a new 3D model for neocognitron, called tricognitron, is presented. One of prominent differences is that tricognitron generates an extra dimension, called altitude dimension from input patterns which can be used to store the important feature characteristics of 2D pattern so as to achieve perfect scale and shift (position) invariance recognition. Based on the altitude dimension a 2D pattern can be represented by a 3D figure pattern generated by tricognitron. A second salient difference of tricognitron is that tricognitron bases its the recognition on feature vectors specified in the altitude dimension rather than training patterns as does neocognitron. A third significant difference is that tricognitron employs only 2-layer modules plus one extra 3D figure layer between C ₁ layer and S₂ layer to cut down 60% of the total number of neurons required for neocognitron. The results are very encouraging and have shown a great promise in pattern recognition