Neural gradients are near-lognormal: improved quantized and sparse training
This work addresses the bottleneck of gradient propagation in training for machine learning practitioners, offering practical improvements in efficiency.
The paper tackled the problem of accelerating neural network training by reducing the computational and memory burdens of neural gradients, achieving state-of-the-art results on ImageNet with 6-bit floating-point quantization and up to 85% gradient sparsity without accuracy degradation.
While training can mostly be accelerated by reducing the time needed to propagate neural gradients back throughout the model, most previous works focus on the quantization/pruning of weights and activations. These methods are often not applicable to neural gradients, which have very different statistical properties. Distinguished from weights and activations, we find that the distribution of neural gradients is approximately lognormal. Considering this, we suggest two closed-form analytical methods to reduce the computational and memory burdens of neural gradients. The first method optimizes the floating-point format and scale of the gradients. The second method accurately sets sparsity thresholds for gradient pruning. Each method achieves state-of-the-art results on ImageNet. To the best of our knowledge, this paper is the first to (1) quantize the gradients to 6-bit floating-point formats, or (2) achieve up to 85% gradient sparsity -- in each case without accuracy degradation. Reference implementation accompanies the paper.