Recognition of pattern position and shape by population vector in spatial spreading associative neural network

Abstract

In the brain, both spatial and object (shape) recognition systems are considered to work in parallel and cooperatively. In this paper, a spatial spreading associative neural net (SSANN) that has both spatial and shape recognition systems is developed. The basic learning algorithm is generalized inverse learning. The characteristics of the SSANN are the incorporation of the spatial spreading neural layer (SSNL) and the use of the population vector method. The SSNL spreads the input pattern by a positional (Gaussian) weight function which has similar tuning characteristics to the directional discrimination neuron found in the parietal cortex. The spread pattern is then associated with the directional and shape memory neurons by generalized inverse learning. For the recognition of the object position in the pattern, a population vector is defined as an ensemble of characteristic vectors of directional memory neurons. The direction of the population vector recognizes the object positions in the input pattern, irrespective of its shape. The shape memory neurons recognize the object shape in the input pattern, irrespective of its position. A non-spreading associative neural net that does not have a SSNL cannot recognize an object's position and shape unless it is located in the memorized position. Thus, the SSNL is critical for correct recognition performance. In consequence, the SSANN achieves both shape-invariant positional recognition and position-invariant shape recognition of the object at the same time, and provides a new tool for pattern recognition.