BeFit: Shifting Complexity from Weights to Bernstein Activations for Resource-Constrained Sensing

Impact 7 arXiv cs.AI

Summary

  • Researchers introduce BeFit, a framework that shifts model capacity from weights to learnable Bernstein polynomial activations for resource-constrained sensing.
  • BeFit uses hardware-aware neural architecture search, knowledge distillation, and weight projection to create multiplication-free convolutional neural networks.
  • On a Cortex-M4 microcontroller, projected BeCNNs achieve a 2.2x speedup while maintaining accuracy comparable to baseline models.
  • The approach also extends to transformers and MLPs, staying within 1.56 percentage points of float32 teacher models on average.

Why it matters

BeFit enables efficient deployment of neural networks on resource-constrained microcontrollers and hardware without native multipliers by shifting computational complexity from weights to activations.

Who should care