Regular and fast chaotic neural network learning for single and translation invariant pattern recognition

In this work the systematic investigations of the two-layer and the three-layer perceptron fast learning process, depending on the number of patterns, its activity, the slope of the neuron response function, the learning rate and the inertial coefficients, were carried out. It is shown that the regular neural network learning is substituted for chaotic learning when we increase the coefficients mentioned above. Such increasing of the learning rate coefficient, the inertial coefficient or the slope of the neuron response function leads to the improvement of the neural network convergence in spite of the output error oscillations. The optimal learning parameters of the two-layer and the three-layer neural networks have been obtained. The study of the learning processes for the translation invariant pattern recognition was also carried out