XNOR-Net: scaled binary-filter and binary-input convolution approximations

What changed

Approximated convolution weights and inputs using signs and real scaling factors, combining XNOR/bitcount operations with binary-network training for ImageNet-scale classifiers.

Scientific scope

Approximated convolution weights and inputs using signs and real scaling factors, combining XNOR/bitcount operations with binary-network training for ImageNet-scale classifiers. Inherits BinaryConnect/BinaryNet and real-valued weight accumulators. Reported32×filter storage and58×convolution operation estimates are not universal whole-model speedups; scaling/nonbinary boundary layers have costs. The sign/mean-absolute solution solves the stated filterL2 approximation, not global network-loss optimization.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits BinaryConnect/BinaryNet and real-valued weight accumulators. Reported32×filter storage and58×convolution operation estimates are not universal whole-model speedups; scaling/nonbinary boundary layers have costs. The sign/mean-absolute solution solves the stated filterL2 approximation, not global network-loss optimization.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction