Research map / Architectures and efficiency / RNNs and normalization
GACT: Activation Compressed Training for Generic Network Architectures
Xiaoxuan Liu, Lianmin Zheng, Dequan Wang, Yukuo Cen, Weize Chen, Xu Han, Jianfei Chen, Zhiyuan Liu, Jie Tang, Joey Gonzalez, Michael Mahoney, Alvin Cheung
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 |
|---|---|---|---|
| Use adaptive per-tensor bit allocation rather than uniform 4-bit quantization (ablation) | GACT on graph neural networks · Flickr, GAT | accuracy: 35.24 ± 11.90 -> 52.26 ± 0.31 | Improved |
| Use adaptive per-tensor bit allocation rather than uniform 4-bit quantization (ablation) | GACT on Bert-large · MNLI | accuracy: 85.98 ± 0.16 -> 86.61 ± 0.11 | Improved |
| Reduce adaptive average precision from 4 bits to 2 bits (ablation) | GACT · ImageNet classification and COCO detection | accuracy: 68.77 -> 68.49; 76.96 -> 76.13; 80.92 -> 77.91; 37.0 -> 36.1; 36.3 -> Div. | Mixed |
| Reduce adaptive average precision from 4 bits to 2 bits (ablation) | GACT on Bert-large · MNLI | accuracy: 86.61 ± 0.11 -> 84.24 ± 0.74 | Hurt |
| Reduce adaptive average precision from 4 bits to 2 bits (ablation) | GACT on graph models · Flickr | accuracy: 51.08 ± 0.18 -> 51.20 ± 0.18; 52.26 ± 0.31 -> 51.62 ± 1.19; 52.31 ± 0.16 -> 52.23 ± 0.15 | Mixed |
| Add swapping of compressed activations (ablation) | GACT on Bert-large-cased · Bert-large-cased | Speed (sequence/s): 6.02 -> 12.95 | Improved |
| Add prefetching to GACT swap (ablation) | GACT swap on Bert-large-cased · Bert-large-cased | Speed (sequence/s): 12.95 -> 14.02 | Improved |
| Add efficient self-attention to CB1 (ablation) | GACT with gradient checkpointing (CB1) · Bert-large batch size 288 | Peak Mem. (MB): 8251 -> 8102 | Improved |
| Combine GACT with gradient checkpointing (ablation) | Bert-large-cased · SST-2 | accuracy: 93.58 -> 93.81 | No clear effect |
| Combine GACT with gradient checkpointing (ablation) | Bert-large-cased · QNLI | accuracy: 92.42 -> 92.26 | No clear effect |
Closest papers by meaning
- AC-GC: Lossy Activation Compression with Guaranteed Convergence (NeurIPS 2021 · similarity 0.88)
- EXACT: Scalable Graph Neural Networks Training via Extreme Activation Compression (ICLR 2022 · similarity 0.88)
- ActNN: Reducing Training Memory Footprint via 2-Bit Activation Compressed Training (ICML 2021 · similarity 0.87)
- N2N learning: Network to Network Compression via Policy Gradient Reinforcement Learning (ICLR 2018 · similarity 0.87)
- Subspace Networks: Scaling Decentralized Training with Communication-Efficient Model Parallelism (NeurIPS 2025 · similarity 0.85)
- 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.
