ConvReflex: Efficient Ultra-Low-Power CNN Inference via Clamping Prediction

ConvReflex enables efficient convolutional neural network (CNN) inference on ultra-low-power devices by predicting clamped (saturating) activations and avoiding their full computation. Skipping work that does not affect the output reduces energy and latency for embedded inference without sacrificing accuracy, making CNN inference practical on severely energy-constrained microcontrollers and sensing platforms.

Recommended citation: S. Li, L. Mottola, Yuan Yao and S. Kaxiras, "ConvReflex: Efficient Ultra-Low-Power CNN Inference via Clamping Prediction," ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems (SenSys), Saint-Malo, France, 2026.
Download Paper