Jesse Levine

Research overview

Jesse researches efficient activation-function implementations in FPGAs. The work connects numerical approximation with hardware design, including the tradeoffs among accuracy, throughput, latency, table structure, arithmetic precision, and FPGA resource use.

The broader MOSAIC program treats nonlinear blocks as part of the inference system rather than isolated mathematical functions. Approximation quality must therefore be judged by the downstream observable the system is trying to protect.

Biomedical and measurement background

Jesse holds a BEng in Biomedical and Electrical Engineering. His undergraduate work includes:

• Raman spectroscopy and machine learning for bacterial classification
• ECG implementation of the Pan–Tompkins detection pipeline in analog hardware
• EMG acquisition and filtering for an interactive game
• Adaptive human-computer and musical interfaces
• Embedded systems and Bluetooth Low Energy logistics

This background joins sensors, analog front ends, filtering, classification, user variability, and experimental validation.

Selected current output

Numerically Stable Softmax at the Edge Using a RISC-V Accelerator for IoT Inference, with Leonard MacEachern, was submitted to IEEE Embedded Systems Letters in June 2026.

Current direction

The current work develops efficient nonlinear-function hardware and the evidence needed to choose approximation structures and precision for a real inference workload.