Improved Projection Learning for Lower Dimensional Feature Maps

Herein we propose an improved method for learning low-rank The requirement to repeatedly move large feature maps offand projections which can be incorporated into pre-trained CNNs on-chip during inference with convolutional neural networks to reduce their maximal memory requirements. So doing, this (CNNs) imposes high costs in terms of both energy approach seeks to both reduce the memory requirements on and time. In this work we explore an improved method for a device, and ideally to eliminate off-chip memory access compressing all feature maps of pre-trained CNNs to below a mid-forward-pass, which can dominate power usage [3, 4], specified limit. This is done by means of learned projections a goal which is would enable lower-power, edge-device deployed trained via end-to-end finetuning, which can then be folded deep networks.