Neocognitron: A Survey of a Classical Hybrid Neural Net work Model
M. Kukaÿcka · 2011
The N eocognitron neural network was introduced in 1980, and since then, it has developed from a model of brain's visual cortex into an effective pattern recognition tool. The model has many noteworthy properties, among others the ability to learn and classify visual patterns without any need for data preprocessing and the major use of self-organization in its learning algorithm. During its development, the model has undergone many modifications and extensions, in some cases gaining interesting abilities, such as the ability to restore damaged patterns. This article provides a survey of the Neocognitron's basic functionality and the ideas of its most interesting modifications. The model has undergone a long development, starting in 1980 with a basic version of the network and continuing in a large number of improvements and modifications, which are described in numerous articles. The network was extended in various manners, with neural circuits being added to provide new functionality, or removed, when another improvement made them redundant. This article aims to provide a description of the basic Neocognitron model together with an overview of the most significant modifications and their effect on the network's performance. The ideas used in the construction of the Neocognitron can be utilized in other models, their influence is found in some of the most successful pattern recognition models, such as Yann LeCun's LeNet (LeCun et al., 1998).