Research On Accelerating Convolutional Neural Networks Via Eliminating Weight And Feature Redundancy

In recent years,deep convolutional neural networks(CNNs)have achieved breakthrough in tasks such as localization,retrieval,detection and recognition of visual content.Through learningable convolution filters,neurons with non-linear output cells,deeper models and large-scale samples,convolution neural network have achieved enough performence to be widely deployed in the server and mobile terminals.However,Convolutional neural networks' performance gain is based on the growing computational complexity.It means that deploying equal amounts of service requires much more hardware than other algorithm in industrial environment.And the expensive training time cost also limits the efficiency in research.Therefore,a convolution neural network accelerating algorithm with low degree of coupling with specific hardware environment is necessary for industrial deploying and researching activities.In this paper,based on the analyzing of convolutional neural network models and the time cost of units in them,a feasible convolutional neural networks accelerating algorithm via eliminating both spatial and channel redundancy is proposed.The innovation of this paper are mainly reflected in the following aspects:1,Research on a new algorithm with combination of feature and parameter compression&Im2col and Pooling optimization algorithm(1)For feature geometry space,a mask-based algorithm is deplopy to eliminate spatial redundancy to achieve acceleration.For the weights channel space,a parameter compression algorithm based on tensor decomposition is deployed.And by combining two defferent methods above,a accelerating algorithm with higher acceleration performance and lower loss is proposed.(2)A Im2col and Pooling layer accelerating algorithm with no performance loss is proposed for particular step size of sliding windows.2,Research on an effective compound accelerating strategy(1)In this paper,the accelerating ratio and performance loss of different kinds of accelerating algorithms is systematically analyzed and the conclusions are verified experimentally.(2)A training optimization algorithm via Dark Knowledge is deployed to recover the performance loss which is brought in during accelerating phase.(3)On the basis of(1)(2),an effective compound accelerating strategy is proposed.