Neocognitron of a new version: handwritten digit recognition

The author (1988) and Fukushima and Miyake (1982) proposed a neural network model, neocognitron, for robust visual pattern recognition. This paper proposes an improved version of the neocognitron and demonstrates its ability using a large database of handwritten digits (ETL-1). To improve the recognition rate of the neocognitron, several modifications have been applied, such as: the inhibitory surround in the connections from S-cells to C-cells, contrast-extracting layer between input and edge-extracting layers, self-organization of line-extracting cells, supervised competitive learning at the highest stage, and so on. These modifications allowed the removal of accessory circuits that were appended to the previous versions, resulting in an improvement of recognition rate as well as simplification of the network architecture. The recognition rate varies depending on the number of training patterns. When we used 3000 digits (300 patterns for each digit) for the learning for example, the recognition rate was 98.5% for a blind test set (3000 digits), and 100% for the training set