Once-for-All Network and AutoML Techniques on FPGA

Designing neural networks for specialized hardware architectures is a trending topic in recent years, especially on edge devices. In this talk, you will see how to search and deploy efficient models for Xilinx’s popular AI platform, Vitis AI at low cost with the help of a Once-for-All (OFA) network. OFA is an AutoML approach that can automatically design efficient and compact neural network architectures fitting the FPGA’s resource, improving the FPGA utilization by 50% compared to conventional mobile neural network architectures. OFA is the award-winning solution in the IEEE Low Power Computer Vision Challenge, FPGA track.