AI Paper Atlas

Research map / Optimization / Optimizers

AC-GC: Lossy Activation Compression with Guaranteed Convergence

R David Evans, Tor Aamodt

NeurIPS 2021 · Optimization › Optimizers · approach: Constrained and non-backprop training · 4 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
replace JPEG-ACT heuristic hyperparameters with AC-GC error bounds (AutoJPEGZ)JPEG compression · Average over all datasetsaccuracy change and compression: −1.0 ( 7.4×) -> −0.6 (10.5×)Improved
use AC-GC bounds with CuSZ instead of grid search over fixpointfixpoint compression · Average over all datasetscompression ratio: 7.8× -> 15.1×Improved
AutoCuSZ vs. JPEG-ACT compressionJPEG-ACT · Average over all datasetscompression ratio: 7.4× -> 15.1×Improved
AutoQuantZ instead of JPEG-ACT on ImageNetJPEG-ACT · ImageNet RN50top-1 accuracy: 71.6 -> 72.7Improved
AutoCuSZ instead of EBC (hand-tuned CuSZ)EBC · ImageNet RN50accuracy change and compression: −0.9 -> −0.2±0.2Improved
AutoCuSZ vs. ActNN-L3ActNN-L3 · ImageNet ResNetscompression ratio: 12× -> 10.1×Hurt
compress activations with AutoQuantZ instead of uncompressedbaseline RN50 · ImageNet RN50top-1 accuracy: 72.3 -> 72.7No clear effect
AutoCuSZ on ImageNetbaseline RN50 · ImageNet RN50top-1 accuracy: 72.3 -> 72.5Hurt
increase e^2 above 0.5 (ablation)AC-GC · all datasetsaccuracyHurt
use mean instead of max summary of statistics (ablation)AutoQuant · MNIST/LeNettraining lossNo clear effect
vary recalculation interval (ablation)AutoQuant · MNIST/LeNettraining lossNo clear effect
decrease e (increase bitwidth) (ablation)AutoQuant · MNIST/LeNetlossHurt

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