DIVISION: Memory Efficient Training via Dual Activation Precision

TL;DR

DIVISION introduces a memory-efficient DNN training method that compresses high-frequency activation components, reducing memory use over 10x while maintaining accuracy and throughput.

Contribution

The paper proposes a novel activation compression technique that separates low- and high-frequency components, simplifying and improving memory efficiency during training.

Findings

Achieves over 10x activation map compression.

Maintains competitive model accuracy.

Offers better performance than existing methods.

Abstract

Activation compressed training provides a solution towards reducing the memory cost of training deep neural networks~(DNNs). However, state-of-the-art work combines a search of quantization bit-width with the training, which makes the procedure complicated and less transparent. To this end, we propose a simple and effective method to compress DNN training. Our method is motivated by an instructive observation: DNN backward propagation mainly utilizes the low-frequency component (LFC) of the activation maps, while the majority of memory is for caching the high-frequency component (HFC) during the training. This indicates the HFC of activation maps is highly redundant and compressible during DNN training, which inspires our proposed Dual Activation Precision (DIVISION). During the training, DIVISION preserves the high-precision copy of LFC and compresses the HFC into a light-weight copy with low numerical precision. This can significantly reduce the memory cost without negatively affecting the precision of backward propagation such that DIVISION maintains competitive model accuracy. Experiment results show DIVISION has better comprehensive performance than state-of-the-art methods, including over 10x compression of activation maps and competitive training throughput, without loss of model accuracy.