Research map / Optimization / Optimizers
AC-GC: Lossy Activation Compression with Guaranteed Convergence
R David Evans, Tor Aamodt
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace JPEG-ACT heuristic hyperparameters with AC-GC error bounds (AutoJPEGZ) | JPEG compression · Average over all datasets | accuracy change and compression: −1.0 ( 7.4×) -> −0.6 (10.5×) | Improved |
| use AC-GC bounds with CuSZ instead of grid search over fixpoint | fixpoint compression · Average over all datasets | compression ratio: 7.8× -> 15.1× | Improved |
| AutoCuSZ vs. JPEG-ACT compression | JPEG-ACT · Average over all datasets | compression ratio: 7.4× -> 15.1× | Improved |
| AutoQuantZ instead of JPEG-ACT on ImageNet | JPEG-ACT · ImageNet RN50 | top-1 accuracy: 71.6 -> 72.7 | Improved |
| AutoCuSZ instead of EBC (hand-tuned CuSZ) | EBC · ImageNet RN50 | accuracy change and compression: −0.9 -> −0.2±0.2 | Improved |
| AutoCuSZ vs. ActNN-L3 | ActNN-L3 · ImageNet ResNets | compression ratio: 12× -> 10.1× | Hurt |
| compress activations with AutoQuantZ instead of uncompressed | baseline RN50 · ImageNet RN50 | top-1 accuracy: 72.3 -> 72.7 | No clear effect |
| AutoCuSZ on ImageNet | baseline RN50 · ImageNet RN50 | top-1 accuracy: 72.3 -> 72.5 | Hurt |
| increase e^2 above 0.5 (ablation) | AC-GC · all datasets | accuracy | Hurt |
| use mean instead of max summary of statistics (ablation) | AutoQuant · MNIST/LeNet | training loss | No clear effect |
| vary recalculation interval (ablation) | AutoQuant · MNIST/LeNet | training loss | No clear effect |
| decrease e (increase bitwidth) (ablation) | AutoQuant · MNIST/LeNet | loss | Hurt |
Closest papers by meaning
- GACT: Activation Compressed Training for Generic Network Architectures (ICML 2022 · similarity 0.88)
- Subspace Networks: Scaling Decentralized Training with Communication-Efficient Model Parallelism (NeurIPS 2025 · similarity 0.88)
- AC/DC: Alternating Compressed/DeCompressed Training of Deep Neural Networks (NeurIPS 2021 · similarity 0.88)
- Compressing Neural Networks: Towards Determining the Optimal Layer-wise Decomposition (NeurIPS 2021 · similarity 0.87)
- Error Feedback Can Accurately Compress Preconditioners (ICML 2024 · similarity 0.86)
- Auto-Compressing Networks (NeurIPS 2025 · similarity 0.85)
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.
