Combining Arithmetic Approximation Techniques for Improved CNN Circuit Design
George Lentaris, George Chatzitsompanis, Vasileios Leon, Kiamal Pekmestzi, Dimitrios Soudris
Abstract
Convolutional Neural Networks (CNN) plea for improved design at circuit level. Instead of compression, we adopt here arithmetic approximation techniques supporting seamlessly the original structure and arithmetic type of a given network. We focus on the convolution and explore various approximations towards refining the resources and/or throughput of any CNN implementation. We develop hybrid high radix multipliers, block floating point arithmetic, as well as parallel architectures and Winograd convolutions to further enhance our design. Results show half cost or even quadruple FPS, depending on ASIC/FPGA technology, with only negligible 0.1-0.4% loss of CNN accuracy.