Search academic texts

Information × Registration Number 0224U031594, 0121U109207 , R & D reports Title Methods and algorithms of optimization of pattern recognition based on machine learning methods popup.stage_title Head Shaldenko Oleksii V., Registration Date 21-05-2024 Organization National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" popup.description2 ABSTRACT The total volume of the report contains 78 pages, 30 figures, the list of used sources contains 53 names. The method for optimizing neural networks for face detection has been improved. This method allows you to exclude important layers in the network from the evaluation, while keeping the number of excluded weights unchanged. Thanks to this, the accuracy of the resulting network increases by 0.7% on the validation sample. The pre-training pruning method has been improved for use on transformer-type models by including attention ratings in the calculation of the weight importance criterion. This modification made it possible to train a network with 34% more accuracy compared to the original method. Experiments on training and accuracy assessment were conducted to confirm the effectiveness of the developed method on a model for solving computer vision problems. Studies have shown that currently the evaluation of pruning methods was carried out on typical architectures of neural network models. When investigating the network for detecting faces, and applying the pruning method before training, it was found that the network can maintain a high level of accuracy, while significantly reducing the number of parameters involved in training. This speeds up the learning and makes it easier to use the models later, as the twisted CPUs and GPUs optimize calculations involving sparse matrices. Keywords: NEURAL NETWORKS, COMPUTER VISION, OPTIMIZATION, PRUNING, TRANSFORMERS.  Product Description popup.authors Zdor Kostiantyn A. Artem V. Melnychenko popup.nrat_date 2024-05-21 Close