Area-efficient Binary and Ternary CNN Accelerator using Random-forest-based Approximation

Abstract

In recent years, the demand for faster inference of convolutional neural networks with a smaller and low-power accelerator is increasing to realize low-latency control of robots and reduce network load. In this paper, we propose a random-forest-based approximation layer unit (RFA-LU) for binary and ternary CNNs to realize faster inference. This unit introduces a novel technique predicting output feature maps using random forest models instead of directly calculating multiply-accumulate (MAC) operations. We demonstrate that the proposed RFA-LU can reduce the number of adaptive logic modules (ALMs) by 56.2% (61.3%) and the number of registers by 85.3% (84.9%) compared with conventional binary (ternary) CNN circuits of the same performance on an Intel Cyclone V SX FPGA.