Skip to main content
Back to timeline
arXivSource publication:

BeFit moves model capacity from weights to Bernstein activations: projected convolutions drop multiplication and land within 1.0 pp of the MLPerf Tiny reference

Synopsis

On microcontrollers with a few hundred kilobytes of memory and no hardware multiplier, the multiplication cost of neural inference has been a deployment bottleneck. BeFit shifts model capacity from weights to learnable Bernstein polynomial activations: a hardware-aware search and knowledge distillation produce a compact network, then convolutional weights are projected to multiplication-free codes and frozen while the activation curves recover the lost accuracy. On a TI CC3220SF, projected BeCNNs lose 0.8 pp on average versus 2.4 pp for ReLU networks; against the MLPerf Tiny DS-CNN reference, the unprojected BeCNN is 1.0 pp more accurate, 2.2 lower in latency and 6.4 lower in flash, while the projected model stays within 1.0 pp of the reference accuracy and runs 5.9 faster.

Source-provided article image: BeFit: Shifting Complexity from Weights to Bernstein Activations for Resource-Constrained Sensing
arXiv

Interpretation

BeFit removes multiplications from convolutional inference on microcontrollers: convolutional weights are projected into codes that need only additions and shifts, and each activation becomes a single table lookup. Earlier microcontroller deployments either kept weights that must be multiplied or restricted weights to ternary or powers of two while keeping fixed activations, so the restricted weights had to absorb the accuracy loss alone. On a TI CC3220SF (Cortex-M4), weight projection speeds BeCNNs up by 2.2 on average while losing 0.8 pp, compared with 2.4 pp for ReLU networks of the same architecture.

The approach puts expressiveness into the activations: each filter or layer carries a learnable Bernstein polynomial curve, and because the curve is evaluated on a fixed interval it compiles offline into a fixed-size lookup table. A Bernstein polynomial is a convex combination of its coefficients over a bounded domain, so once the input is normalized to that domain the coefficients directly control the curve shape, and inference costs one table read regardless of degree. Across six sensing tasks the final models stay within 1.56 pp of their float32 teachers on average; in emulation on a RISC-V core without a hardware multiplier, the instruction reduction relative to the DS-CNN grows from 5.2 on the Cortex-M4 to 147.

Perspective

The demonstrated scope is end-to-end deployment on a TI CC3220SF board across six sensing tasks (three HAR datasets, Speech Commands keyword spotting and UJIIndoorLoc indoor localization); instruction counts come from QEMU and are checked against single-step counts on the board, while the multiplier-less RISC-V results come from emulation rather than silicon.

The int8 classifier still multiplies, and in the keyword transformers the attention products and softmax also keep their multiplications, making up 20–34% of the unprojected model's multiply-accumulates and limiting the speedup; energy per inference has not been measured on the board, and behavior on multiplier-less RISC-V silicon and Cortex-M cores with vector extensions remains to be tested.