Semi-Structured Sparsity Using Dynamic Percentile

Le résumé fourni par la source

The computational efficiency of deep neural networks is limited by hardware accelerators and sparsity integration. Leveraging sparsity can result in significant computational savings by reducing the number of operations and memory accesses required. In recent years, N:M sparsity, where N out of every M consecutive elements are zero, has emerged as a promising approach for optimizing neural networks. This semi-structured sparsity pattern facilitates high performance at substantial sparsity levels while enabling significant speedups on specialized hardware. However, existing methods often assume the minimum N values to be unimportant, regardless of their relative values. In this paper, we propose p:M, a novel framework that leverages dynamic percentiles in semi-structured sparsity, optimizing model accuracy for higher sparsities. Our approach reformulates the sparsity constraint, introducing dynamic percentiles for masking weights. This dynamic percentile process enables efficient pruning during training without additional computational overhead. Experimental results demonstrate that our method consistently improves upon traditional N:M sparsity techniques across various neural network architectures and datasets, achieving superior performance with reduced computational cost.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.