Machine-Learning-Based Perception on a Tiny, Low-Power FPGA

We present a set of machine-learning-based perception solutions that we implemented on a tiny (5.4 mm2 package), low-power FPGA. These solutions include hand gesture classification, human detection and counting, local face identification, location feature extraction, front-facing human detection and shoulder surfing detection, among others. We describe our compact processing engine structure that fits into fewer than 5K FPGA look-up tables, yet can support networks of various sizes. We also describe how we selected networks and the optimizations we used to make them suitable for low-power and low-cost edge applications. Last but not least, we also describe how we leverage the on-the-fly self-reconfiguration capability of FPGAs to enable running a sequence of processing engines and neural networks in a single FPGA.

There will be a live Q&A session with the presenter immediately following the presentation.