Research Of Comparison Deep Learning Libraries To The Problem Of Classification Of Handwritten Digits

At present,machine learning is an actively developing field of research.This is connected both with the possibility of faster,easier and cheaper to collect and process data,and with the development of methods for identifying laws from these data,according to which physical,biological,economic and other processes take place.In some tasks,when such a law is difficult to determine,use deep learning.In-depth training examines methods for modeling high-level abstractions in data using a variety of sequential non-linear transformations,which,as a rule,are represented as artificial neural networks.Today,neural networks are successfully used to solve problems such as forecasting,pattern recognition,data compression,and several others.The purpose of this study is to conduct a comparative analysis of some software tools of deep learning libraries,such as Caffe,Pylearn2,Torch and Theano.This study will be conducted on the example of solving one of the tasks of deep learning,namely the problem of recognizing handwritten numbers.The MNIST handwritten digital image database will be used as the study dataset.In the course of the study,experiments with CPU and GPU will be conducted for two types of neural network topologies: 1)MLP and 2)CNN.The result of the research is the evaluation of each library by six criteria: 1)learning speed,2)classification speed,3)ease of use,4)configuration flexibility,5)functionality,6)accessibility and ease of use of documentation and training materials.In the future,we plan to use one of the libraries to solve the problems of finding people,pedestrians and cars.