NEURAL NETWORKS TRAINING BASED ON RANDOM SEARCH

The paper deals with a state-of-art problem, associated with neural networks training. Training algorithm (with special parallelization procedure) implementing the annealing method is proposed. The training efficiency is demonstrated by the example of a neural network architecture focused on parallel data processing. For the color image compression problem, it is shown that the proposed algorithm significantly outperforms gradient methods in terms of efficiency. The results obtained make it possible to improve the neural networks training quality in general, and can be used to solve a wide class of applied problems.