Memory Bounded Deep Convolutional Networks

TL;DR

This paper explores sparsity-inducing regularizers in CNN training to create memory-efficient models, demonstrating significant reductions in memory use with minimal accuracy loss across multiple datasets.

Contribution

It introduces a regularization technique that encourages sparse connectivity in CNNs, enabling memory and runtime savings during deployment.

Findings

Memory reduction by a factor of four on AlexNet

Effective sparsity with minimal accuracy loss

Applicable to multiple datasets including ImageNet

Abstract

In this work, we investigate the use of sparsity-inducing regularizers during training of Convolution Neural Networks (CNNs). These regularizers encourage that fewer connections in the convolution and fully connected layers take non-zero values and in effect result in sparse connectivity between hidden units in the deep network. This in turn reduces the memory and runtime cost involved in deploying the learned CNNs. We show that training with such regularization can still be performed using stochastic gradient descent implying that it can be used easily in existing codebases. Experimental evaluation of our approach on MNIST, CIFAR, and ImageNet datasets shows that our regularizers can result in dramatic reductions in memory requirements. For instance, when applied on AlexNet, our method can reduce the memory consumption by a factor of four with minimal loss in accuracy.