Metaheuristic based framework for Deep Neural Networks compression via DeepCABAC Arithmetic Coding and Fractional Filter

Fekhr Eddine Keddous, Amir Nakib

Abstract: This paper is dedicated to the deep neural network compression problem. To deal with this issue, we propose a new framework that includes the compression-efficient DeepCABAC arithmetic coding, a neural network preprocessing strategy for parameter reduction, a fractional filter (FF) to reduce the number of kernels in the considered deep CNNs, and the convolution-batch norm fusion to increases the inference speed. The whole framework is, then, optimized via a metaheuristic. The proposed framework demonstrates an impressive compression efficiency of up to 94% on VGG-16 without loss on the classification quality, as measured by top-l and top-5 accuracy metrics. For applications where accuracy is not a primary concern, the coding gains can be further increased.