Research map / Architectures and efficiency / Pruning and efficient tuning
AC/DC: Alternating Compressed/DeCompressed Training of Deep Neural Networks
Alexandra Peste, Eugenia Iofinova, Adrian Vladu, Dan Alistarh
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 alternating dense and sparse training phases | ResNet50 · ImageNet, 80% sparsity | Top-1 accuracy (%): 74.6 \pm 0.06 -> 76.3 \pm 0.1 | Improved |
| Use alternating dense and sparse training phases | ResNet50 · ImageNet, 90% sparsity | Top-1 accuracy (%): 74.76 -> 75.03 \pm 0.1 | Improved |
| Use alternating dense and sparse training phases | ResNet50 · ImageNet, 95% sparsity | Top-1 accuracy (%): 72.12 -> 73.14 \pm 0.2 | Improved |
| Use alternating dense and sparse training phases | ResNet50 · ImageNet, 98% sparsity | Top-1 accuracy (%): 67.06 -> 68.44 \pm 0.09 | Improved |
| Use alternating dense and sparse training phases | MobileNetV1 · ImageNet, 75% sparsity | Top-1 accuracy (%): 70.09 -> 70.3 \pm 0.07 | Improved |
| Use alternating dense and sparse training phases | MobileNetV1 · ImageNet, 90% sparsity | Top-1 accuracy (%): 63.87 -> 66.08 \pm 0.09 | Improved |
| Prune embeddings to 50% sparsity (ablation) | AC/DC, Transformer-XL · WikiText-103, 80% model sparsity | perplexity, sparse: 20.65 -> 20.83 | Hurt |
| Prune embeddings to 50% sparsity (ablation) | AC/DC, Transformer-XL · WikiText-103, 90% model sparsity | perplexity, sparse: 22.32 -> 22.84 | Hurt |
| Shorten dense phases from five to two epochs and omit the longer final decompression phase (ablation) | AC/DC, ResNet50 · ImageNet, 90% global sparsity | Top-1 accuracy (%): 75.03 -> 74.6 | Hurt |
| Use uniform rather than global sparsity (ablation) | AC/DC, ResNet50 · ImageNet, 90% sparsity | Top-1 accuracy (%): 75.03 -> 75.04 | No clear effect |
| Keep the first and last layers dense (ablation) | AC/DC, ResNet50 · ImageNet, 90% global sparsity | Top-1 accuracy (%): 75.03 -> 75.64 | Improved |
| Keep the first layer and depth-wise convolutions dense (ablation) | AC/DC, MobileNetV1 · ImageNet, 90% sparsity | Top-1 accuracy (%): almost 0.5% | Improved |
Later papers that compared with this method
6 later accepted papers report AC/DC as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- AC/DC on a Budget -- Alternating Sparse Phases (NeurIPS 2026 · better on 3 of 4 shared results)
- SequentialAttention++ for Block Sparsification: Differentiable Pruning Meets Combinatorial Optimization (NeurIPS 2024 · better on 17 of 17 shared results)
- Finding Transformer Circuits With Edge Pruning (NeurIPS 2024 · better on 4 of 4 shared results)
- Scaling Laws for Sparsely-Connected Foundation Models (ICLR 2024 · better on 0 of 2 shared results)
- Towards Higher Ranks via Adversarial Weight Pruning (NeurIPS 2023 · better on 10 of 10 shared results)
- PDP: Parameter-free Differentiable Pruning is All You Need (NeurIPS 2023 · better on 3 of 3 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Top-KAST — Top-KAST: Top-K Always Sparse Training (NeurIPS 2020) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AC/DC on a Budget -- Alternating Sparse Phases (NeurIPS 2026 · similarity 0.89)
- Get More at Once: Alternating Sparse Training with Gradient Correction (NeurIPS 2022 · similarity 0.88)
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (ICLR 2024 · similarity 0.87)
- Sparse Cocktail: Every Sparse Pattern Every Sparse Ratio All At Once (ICML 2024 · similarity 0.86)
- Global Minimizers of $\ell^p$-Regularized Objectives Yield the Sparsest ReLU Neural Networks (NeurIPS 2025 · similarity 0.86)
- A Single-Step, Sharpness-Aware Minimization is All You Need to Achieve Efficient and Accurate Sparse Training (NeurIPS 2024 · similarity 0.86)
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.
