Split learning on FPGAs

Abstract

MIT Lincoln Laboratory is developing a software-reconfigurable imaging architecture called ReImagine, the first field programmable imaging array (FPIA), which reflects a broader trend in the increased use of FPGAs in sensor systems in order to reduce size and power consumption without a corresponding loss in performance or flexibility. At the same time, the field of machine learning is diversifying to include distributed deep learning methods like split learning, which can help preserve privacy by avoiding the sharing of raw data and model details. In order to continue to expand the capabilities of architectures like ReImagine's and enable split learning and related techniques to be used in the growing body of FPGA-based sensor systems, we examine the relationship of emerging split learning applications to FPGA-based image processing platforms. We determine that the implementation of split learning methods on FPGAs is feasible, and outline use cases in the areas of health and short timescale physics that demonstrate the usefulness of these implementations to both organizations concerned with privacy-preserving machine learning methods and organizations concerned with the deployment of efficient, flexible, and low-latency sensor systems. We begin by conducting a survey of the modern FPGA landscape in terms of technical attributes, use in sensor systems, security and privacy features, and current machine learning implementations. We also provide an overview of split learning and other distributed deep learning methods. Next, we synthesize an example split learning model in HDL code to demonstrate the feasibility of implementing such a model on an FPGA. Finally, we develop use cases for split learning applications on FPGA-based sensor systems and offer conclusions about the future development of distributed deep learning on heterogeneous processing platforms.