Minimizing Multiplication of Kernel Computation in Convolutional Neural Networks Using Strassen Algorithm

Abstract

Convolution neural networks (CNN) have been widely applied for the computer vision task. However, the success of CNN is limited by the computational complexity of the network, so it is difficult for the model to run the inference process in real time. In this paper, we apply Strassen matrix multiplication to reduce multiplications in convolution operations in CNN, in order to get faster execution for CNN. First, we transform the convolution operation into a matrix multiplication operation using the Toeplitz mapping method, then after that, we apply the Strassen method to these matrices. In the end, we compare the number of arithmetic operations (multiplication and addition) in the convolutional layer using Strassen and the standard algorithm. We apply this algorithm implementation in convolution layers 1 and 3 in LeNet-5 Architecture.