An In-depth Study of Stochastic Backpropagation

TL;DR

This paper investigates Stochastic Backpropagation (SBP) as a memory-efficient method for training deep neural networks, demonstrating significant GPU memory savings with minimal accuracy loss in image tasks.

Contribution

It introduces SBP as an effective stochastic gradient approximation technique that reduces memory and computation during backpropagation in deep learning.

Findings

SBP saves up to 40% GPU memory during training.

SBP causes less than 1% accuracy degradation.

SBP speeds up training with minimal impact on performance.

Abstract

In this paper, we provide an in-depth study of Stochastic Backpropagation (SBP) when training deep neural networks for standard image classification and object detection tasks. During backward propagation, SBP calculates the gradients by only using a subset of feature maps to save the GPU memory and computational cost. We interpret SBP as an efficient way to implement stochastic gradient decent by performing backpropagation dropout, which leads to considerable memory saving and training process speedup, with a minimal impact on the overall model accuracy. We offer some good practices to apply SBP in training image recognition models, which can be adopted in learning a wide range of deep neural networks. Experiments on image classification and object detection show that SBP can save up to 40% of GPU memory with less than 1% accuracy degradation.