Weight Sparseness for a Feature-Map-Split-CNN Toward Low-Cost Embedded FPGAs

Convolutional neural network (CNN) has a high recognition rate in image and are used embedded systems such as smartphones, robots self-driving cars. Low-end FPGAs candidates for platforms because they achieve real-time performance at low cost. However, CNN significant parameters called weights internal data feature maps, which pose challenge memory capacity. To solve these problems, we exploit split-CNN weight sparseness. The reduces the footprint by splitting map into smaller patches allows to be stored FPGA's high-throughput on-chip memory. Weight sparseness computational costs achieves even higher performance. We designed dedicated architecture of sparse buffering scheduling implemented this on PYNQ-Z1 FPGA board with low-end FPGA. An experiment classification using VGG16 shows that our implementation is 3.1 times faster than GPU, 5.4 an existing implementation.