Machine learning in conditions of low data availability

Abstract

Machine learning is the subject of numerous scientific and professional research projects and is an important component of systems used in medicine, banking, computer security, communications and numerous other fields. It is one of the most active areas of research with constant progress and development of new algorithms and approaches as well as improvement of existing methods. The performance of the machine learning model is significantly affected by the dataset used for training, i.e. the quality of the data, the uniform distribution of values and the size of the set. This is a potential problem with machine learning methods that require pre-labelled data, as data acquisition can be extremely complex, expensive and time-consuming. In this case, the classical machine learning model will most likely not perform well. One approach to solve this problem is to apply transfer learning, where the model uses a dataset not only from the target domain but also from other, and ideally related domains. In the work, conditions with lower availability of datasets were simulated, under which the performance of three models was analyzed, one of which was based on a previously trained model. The process of creating training sets is described, and the results of analyzing the three models with different sized sets are presented.