AI Paper Atlas

Research map / Architectures and efficiency / Quantization

CAT: Compression-Aware Training for bandwidth reduction

Chaim Baskin, Brian Chmiel, Evgenii Zheltonozhskii, Ron Banner, Alex M. Bronstein, Avi Mendelson

JMLR 2021 · Architectures and efficiency › Quantization · approach: Post-training weight quantization · 11 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add entropy penalty (lambda 0.05) during training (ablation)CAT · ImageNet val (ResNet-18, 5-bit)memory bits per value: 3.100 -> 1.930 (6-bit compute, lambda 0 vs 0.05)Improved
increase lambda from 0.05 to 0.3 (ablation)CAT · ImageNet val (ResNet-18, 5-bit)top-1 accuracy (%): 69.490 -> 67.700Hurt
replace soft entropy with compressibility loss (ablation)CAT · ImageNet val (ResNet-18)top-1 accuracy (%): 67.86% -> 67.84% (4-bit); 69.49% -> 69.36% (5-bit)No clear effect
increase batch size from 16 to 64 (ablation)CAT · ImageNet val (ResNet-50, 4-bit)top-1 accuracy (%): 73.800 -> 74.5Improved
exponential temperature scheduling for soft entropy (ablation)CAT · ImageNet valtop-1 accuracy (%)No clear effect
vary random seed over multiple runs (ablation)CAT · ImageNet val (ResNet-18, 5-bit)top-1 accuracy (%): 69.122±0.016 over 5 runsNo clear effect
CAT versus prior detection compression on SSD512CAT · PASCAL VOC detectionaccuracy (%): 68.12 -> 64.39 (prior, 8-bit); 80.72 -> 77.49 (CAT, 6-bit)Improved
add CAT 5-bit compute to ResNet-18CAT · ImageNet valcompression ratio: 1 -> 3.33; accuracy 69.70 -> 69.20Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy