Research On Fast Algorithm For Convolution Neural Networks

Convolution neural networks(CNNs)achieves a great success in computer vision fields,such as image classification,video analysis and objection detection.This is because a convolution neural network has a great ability of extracting features.Although convolution neural network performs well in many fields,it's algorithmic complexity and storage consumption become bottleneck that hampers it's development.In some embedded devices that hold limit computational ability and storage space,convolution neural network cannot run fast on them.This leads to a practical difficulty.In this work,we reduce the algorithmic complexity of a 3D CNN to accelerate this model with Winograd's minimal algorithm.Winograd algorithm is suitable for minimal convolution rather than large-scale convolution.We use this algorithm on large-scale images by divide the images into small tiles.With this method we reduce 3D CNN's algorithmic complexity.We benchmark a net model on a GPU platform,resulting in a speed-up by a factor of 1.4× compared with cuDNN,which is commonly used in many current machine learning frameworks.What is more,we analyze the computing complexity of direct convolution and fastFourier-transform-based(FFT-based)convolution.We creatively propose CS-layer,which is equivalent to a combination of a convolution layer and a pooling layer but more effective.Theoretical computing complexity of CS-layer and some other similar operation is demonstrated,revealing an advantage on computation of CS-layer.Also,practical experiments are also performed and the result shows that CS-layer holds a real superiority on run time.