Improved Projection Learning for Lower Dimensional Feature Maps

TL;DR

This paper proposes an improved projection learning method to compress CNN feature maps below a size limit, enabling more efficient on-chip inference by end-to-end finetuning and folding of learned projections.

Contribution

It introduces a learned projection approach for feature map compression and a ceiling compression framework for on-chip inference optimization.

Findings

Achieved significant reduction in feature map size with minimal accuracy loss.

Demonstrated the feasibility of fully on-chip CNN inference with compressed feature maps.

Provided a new framework for future energy-efficient neural network deployment.

Abstract

The requirement to repeatedly move large feature maps off- and on-chip during inference with convolutional neural networks (CNNs) imposes high costs in terms of both energy and time. In this work we explore an improved method for compressing all feature maps of pre-trained CNNs to below a specified limit. This is done by means of learned projections trained via end-to-end finetuning, which can then be folded and fused into the pre-trained network. We also introduce a new `ceiling compression' framework in which evaluate such techniques in view of the future goal of performing inference fully on-chip.