Apparatus for compressing deep learning models, computer-implemented method and medium

Abstract

A computer-implemented method includes: in an AI system comprising a deep learning model with multiple layers having a weight matrix defining the model, determining groups of two or more consecutive layers within the multiple layers; for each determined group of two or more consecutive layers, applying Singular Value Decomposition (SVD) to the weights of one layer in that group; for each layer to which SVD was applied, truncating at least some singular values ​​obtained from the SVD application according to at least one truncation criterion; for each determined group of two or more consecutive layers, selecting multiple output channels in the deep learning model for removal based on at least one weight importance criterion; and for each determined group of two or more consecutive layers, removing one or more input channels and / or one or more output channels based on the selected multiple output channels, such that the common dimension size in the group remains unchanged. Furthermore, an apparatus and a computer program are also provided.

Images