Research on Self-Supervised Learning as a Potential Alternative to Transfer Learning

The transfer learning allows you to solve applied problems even in the absence of representative training samples. However, its use imposes restrictions on changes in the network architecture. They are often needed to reduce the computational complexity of a neural network. At the same time, training the network first on a large dataset, and then on a target one is extremely expensive. The paper explores the method of self- supervised learning of the network on a training sample to improve quality. For the experiments, we used our own dataset for the task of classifying solid household waste. The results are compared with learning from scratch, as well as with the transfer learning. According to the research results, self-supervised learning increases the accuracy of the network by 3.6%. The transfer of training still demonstrates the greatest accuracy. At the same time, self-supervised learning from scratch is inferior by only 1.4%, which allows us to assert the applicability of self- supervised learning methods as an alternative to transfer learning. This is especially important when using neural architecture search methods to reduce network computing in solving applied problems.