Study of Initialization Options for Applications with Automatic Machine Learning

Automatic machine learning systems are becoming more and more in demand with the increasing use of neural networks in various products. And it is necessary to prepare a pipeline for them, in which the neural network may require additional training. This is what this work focused on - examination of various approaches to initializing the weights of neural networks for its use in an automatic garbage sorting facility. During the work, three strategies for initializing weights have been proposed, which seem to be the most efficient. We describe the datasets used to compare initialization strategies; the studies conducted, which contain several ways to change the dataset in the process of modeling the operation of the system and show obtained results, based on which some conclusions are made. The result of the work is recommendations on the use of strategies in specific cases encountered in the development of automated machine learning systems.