AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Overcoming Oscillations in Quantization-Aware Training

Markus Nagel, Marios Fournarakis, Yelysei Bondarenko, Tijmen Blankevoort

ICML 2022 · Architectures and efficiency › Quantization · approach: Post-training weight quantization · 17 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
Re-estimate batch-normalization statistics after trainingMobileNetV2 QAT · ImageNet, 3-bit weightsvalidation accuracy (%): 64.97 -> 69.50Improved
Add oscillation dampeningMobileNetV2 LSQ · ImageNet, 4/4-bit weights/activationsvalidation accuracy (%): 69.5 -> 70.5Improved
Add iterative weight freezingMobileNetV2 LSQ · ImageNet, 4/4-bit weights/activationsvalidation accuracy (%): 69.5 -> 70.6Improved
Add oscillation dampeningMobileNetV2 LSQ · ImageNet, 3/3-bit weights/activationsvalidation accuracy (%): 65.3 -> 67.8Improved
Add iterative weight freezingMobileNetV2 LSQ · ImageNet, 3/3-bit weights/activationsvalidation accuracy (%): 65.3 -> 67.6Improved
Add oscillation dampeningMobileNetV3-Small LSQ · ImageNet, 3/3-bit weights/activationsvalidation accuracy (%): 52.0 -> 59.0Improved
Add iterative weight freezingEfficientNet-Lite LSQ · ImageNet, 4/4-bit weights/activationsvalidation accuracy (%): 72.3 -> 73.5Improved
Use fixed, stronger dampening (λ=10^{-2}) rather than the baseline (ablation)MobileNetV2 with 3-bit weights · ImageNet, post-BNvalidation accuracy (%): 69.50 -> 68.57Hurt
Use a cosine-annealed dampening coefficient instead of no dampening (ablation)MobileNetV2 with 3-bit weights · ImageNet, post-BNvalidation accuracy (%): 69.50 -> 70.37Improved
Set a constant freezing threshold to 0.015 (ablation)MobileNetV2 with 3-bit weights · ImageNet, post-BNvalidation accuracy (%): 69.50 -> 70.13Improved
Anneal the freezing threshold from 0.04 to 0.01 (ablation)MobileNetV2 with 3-bit weights · ImageNet, post-BNvalidation accuracy (%): 69.50 -> 70.33Improved
Use too-low constant freezing threshold (0.01) (ablation)MobileNetV2 with 3-bit weights · ImageNet, post-BNvalidation accuracy (%): 69.50 -> 69.18Hurt

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