Development of Sample-Efficient Algorithms for Training of Neural

The study focuses on the development of algorithms, that require less interactions with training data. At this point of the study, the approach is focused on estimating more accurate influence on loss of potential updates of trainable parameters of a model. Having the knowledge of the influence on loss in terms of each model parameter, the weight update can be more relevant. The approach may enable to maintain appropriate direction of model updates for high values of learning rate, increasing sample efficiency of training.