BeFit Framework Improves Accuracy for Resource-Constrained Sensing
November 19, 2026
BeFit shifts model capacity from weights to learnable Bernstein polynomial activations to enable multiplication-free convolutions on microcontrollers. This approach uses hardware-aware neural architecture search and knowledge distillation to achieve a 2.75 percentage point accuracy increase in human activity recognition tasks.
HOW THIS AFFECTS YOU
●
builderYou can deploy more accurate models on low-power microcontrollers by using Bernstein activations instead of heavy weight matrices.