Reduction of Multiplications in Convolutional Neural Networks
Munawar Ali, Baoqun Yin, Aakash Kunar, Ali Muhammad Sheikh, Hazrat Bilal
Abstract
Convolution neural network (CNN) widely used for the application of computer vision, such applications would be beneficial for CNN if their workload could be reduced. In this paper, we evaluate CNN's arithmetic properties (Matrix Multiplications and Additions) and proposed a Strassen algorithm with the help of Pan's result to minimize their workload of computation. That algorithm needs the least workspace and it has great computational accuracy. Matrix multiplication (MM) is the most fundamental computational operation, and the performance of Matrix multiplication (MM) depends on different factors. Moreover, that is a very flexible and robust structure. Here we are focusing on the number of element-wise multiplications and addition, the effects of this research on modern computations.