Seprox

Abstract

Compression techniques such as quantization and pruning are indispensable for deploying state-of-the-art Deep Neural Networks (DNNs) on resource-constrained edge devices. Quantization is widely used in practice – many commercial platforms already support 8bits, with recent trends towards ultra-low precision (4-bits and below). Pruning, which increases network sparsity (incidence of zero-valued weights), enables compression by storing only the nonzero weights and their indices. Unfortunately, the compression benefits of pruning deteriorate or even vanish in ultra-low precision DNNs. This is due to (i) the unfavorable tradeoff between the number of bits needed to store a weight (which reduces with lower precision) and the number of bits needed to encode an index (which remains unchanged), and (ii) the lower sparsity levels that are achievable at lower precisions.We propose Seprox, a new compression scheme that overcomes the aforementioned challenges by exploiting two key observations about ultra-low precision DNNs. First, with lower precision, fewer weight values are possible, leading to increased incidence of frequently-occurring weights and weight sequences. Second, some weight values occur rarely and can be eliminated by replacing them with similar values. Leveraging these insights, Seprox encodes frequently-occurring weight sequences (as opposed to individual weights) while using the eliminated weight values to encode them, thereby avoiding indexing overheads and achieving higher compression. Additionally, Seprox uses approximation techniques to increase the frequencies of the encoded sequences. Across six ultralow precision DNNs trained on the Cifar10 and ImageNet datasets, Seprox achieves model compressions, energy improvements and speed-ups of up to 35.2%, 14.8% and 18.2% respectively.