Memory Saving Method for Enhanced Convolution of Deep Neural Network

Convolution neural network (CNN) is a typical feedforward neural network, which gains extraordinary performance in image recognition. Features are extracted by convolution layers automatically for classification, but high in computation complexity. As an enhanced convolution algorithm, image to column (im2col) method accelerates the calculation with redundant memory overhead. In this work, we present Memory Saving Method (MSM) to improve convolution efficiency with lower memory consumption by elements rearrangement of input blocks. Block calculation can be executed both in serial and in parallel for their independence. It is demonstrated by the experimental results that MSM achieves the same acceleration effect yet sparing memory space for approximately two orders of magnitude with no drop of accuracy.