Low-Overhead Early-Stopping Policies for Efficient Random Forests Inference on Microcontrollers
Abstract
Random Forests (RFs) are popular Machine Learning models for edge computing, due to their lightweight nature and high accuracy on several common tasks. Large RFs however, still have significant energy costs, a serious concern for battery-operated ultra-low-power devices. Following the adaptive (or dynamic) inference paradigm, we introduce a hardware-friendly early stopping policy for RF-based classifiers, halting the execution as soon as a sufficient prediction confidence is achieved. We benchmark our approach on three state-of-the-art datasets relative to different embedded classification tasks, and deploy our models on a single core RISC-V microcontroller. We achieve an energy reduction ranging from 18% to more than 91%, with an accuracy drop lower than 0.5%. Additionally, we compare our approach with other early-stopping policies, showing that we outperform them.
