Random Search in Neural Networks Training

The paper deals with a state-of-art applied problem related to the neural networks training. It is shown that, given the expansion of the range of practical problems, gradient methods do not always satisfy the conditions of the subject area, which contributes to the development of alternative training methods. An original training algorithm is proposed that implements the annealing method, for which convergence to the optimal solution is proven. A modified version of the algorithm has been developed that is invariant to the size of the training sample. Experimental studies (using the example of solving problems of image classification and color image compression) confirm the effectiveness of the proposed approach.