Distortion-tolerance curve of the neocognitron with various structures, including pyramid

The neocognitron is a neural model which can recognize a pattern even when it is shifted in position, changed in size, or distorted. However, the complexity of its structure and its algorithm makes it difficult to tune its parameters and to specify the detail of its structure. The characteristics of this model are measured as the relationship between the degree of distortion of the input and the tolerance of the model. Computational experiments show that the pyramid has a suitable structure for implementing a neocognitron, provided that the input has a relatively low spatial frequency.<>