Multi-threaded learning control mechanism for neural networks

Neural networks are applicable in many solutions for classification, prediction, control, etc. The variety of purposes is growing but with each new application the expectations are higher. We want neural networks to be more precise independently of the input data. Efficiency of the processing in a large manner depends on the training algorithm. Basically this procedure is based on the random selection of weights in which neurons connections are burdened. During training process we implement a method which involves modification of the weights to minimize the response error of the entire structure. Training continues until the minimum error value is reached however in general the smaller it is, the time of weight modification is longer. Another problem is that training with the same set of data can cause different training times depending on the initial weight selection. To overcome arising problems we need a method that will boost the procedure and support final precision. In this article, we propose the use of multi-threading mechanism to minimize training time by rejecting unnecessary weights selection. In the mechanism we use a multi-core solution to select the best weights between all parallel trained networks. Proposed solution was tested for three types of neural networks (classic, sparking and convolutional) using sample classification problems. The results have shown positive aspects of the proposed idea: shorter training time and better efficiency in various tasks. (C) 2018 Elsevier B.V. All rights reserved.