AIP: Saving the DRAM Access Energy of CNNs Using Approximate Inner Products

Abstract

In this work, we propose AIP (Approximate Inner Product), which approximates the inner products of CNNs’ fully-connected (FC) layers by using only a small fraction (e.g., one-sixteenth) of parameters. We observe that FC layers possess several characteristics that naturally fit AIP: the dropout training strategy, rectified linear units (ReLUs), and top-n operator. Experimental results show that 48% of DRAM access energy can be reduced at the cost of only 2% of top-5 accuracy loss (for VGG-f).