Deep information-extreme machine learning for recognition of electromyographic biosignals

The aim of the study is to improve the accuracy of machine learning of the electromyographic signal (EMG) recognition system for controlling a hand prosthesis with a non-invasive biosignal reading system.The information-extreme machine learning method is proposed, which is developed within the framework of the func-tional approach to modelling cognitive processes of building and making classification decisions by natural intelligence. This makes the method flexible when retraining the system in the case of expanding the alphabet of recognition classes. To solve the multidimensionality problem of the recognition class alphabet, a hierarchical information-extreme machine learning algorithm based on the decursive data structure was developed and implemented. This data structure allows switching from multi-class machine learning to two-class machine learning for each stratum of the binary decursive tree when the number of recognition classes increases. Construction of highly accurate decision rules for the recognition classes of each stratum of the binary decomposition tree is carried out by optimising the machine learning parameters, the number of which determines the level of machine learning depth. As the criterion for optimising machine learning parameters, the authors use the Kullback information measure modified by the authors in the form of a functional of the accuracy characteristics.Due to computer modelling, error-free decision rules for the alphabet with eight recognition classes, which characterise the EMG signals of the corresponding hand and finger movements, have been built using the training matrix.Conclusions: It has been experimentally proved that when the number of recognition classes exceeds two, it is necessary to switch from a linear data structure to a hierarchical one in the form of a binary decursive tree.