Research map / Architectures and efficiency / RNNs and normalization
Top-KAST: Top-K Always Sparse Training
Siddhant Jayakumar, Razvan Pascanu, Jack Rae, Simon Osindero, Erich Elsen
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 |
|---|---|---|---|
| select extra backward units randomly instead of by magnitude (ablation) | Top-KAST · ImageNet ResNet-50, fwd 90% bwd 80% | top-1 accuracy: 73.03 -> 74.76 | Improved |
| select extra backward units by magnitude (vs random) (ablation) | Top-KAST · ImageNet ResNet-50, fwd 95% bwd 90% | top-1 accuracy: 68.48 -> 70.42 | Improved |
| allow exploration updates to extra units for first 5000 steps (ablation) | Top-KAST · ImageNet ResNet-50, fwd 90% bwd 0% | top-1 accuracy: 68.26 -> 72.05 | Improved |
| allow exploration updates until step 16000 (ablation) | Top-KAST · ImageNet ResNet-50, fwd 90% bwd 0% | top-1 accuracy: 68.26 -> 74.14 | Improved |
| allow exploration for the full 32000 steps vs 16000 (ablation) | Top-KAST · ImageNet ResNet-50, fwd 90% bwd 0% | top-1 accuracy: 74.14 -> 74.65 | Improved |
| compute top-K every 100 steps instead of every step (ablation) | Top-KAST · ImageNet ResNet-50, fwd 80% bwd 50% | top-1 accuracy: 75.03 -> 75.14 | No clear effect |
| compute top-K every 100 steps instead of every step (ablation) | Top-KAST · ImageNet ResNet-50, fwd 95% bwd 90% | top-1 accuracy: 70.42 -> 70.38 | No clear effect |
| use sparse backward pass (bwd 80%) instead of dense backward | Top-KAST · WikiText-103 | perplexity: 19.8 -> 21.3 | Hurt |
| increase forward sparsity to 95% (bwd 90%) | Top-KAST · WikiText-103 | perplexity: 19.8 -> 32.2 | Hurt |
| Top-KAST sparse model vs smaller dense model | Top-KAST 80% sparse 57M · WikiText-103 | perplexity: 21.5 -> 19.8 | Improved |
| use Top-KAST with dense backward vs magnitude pruning | Top-KAST · enwik8 69M model, fwd 80% | BPC: 1.02 -> 1.03 | Hurt |
| Top-KAST sparse transformer vs dense | Top-KAST (80%, 0%) 55M · enwik8 | BPC: 0.99 -> 1.00 | No clear effect |
Later papers that compared with this method
4 later accepted papers report Top-KAST as a baseline in their results tables; 4 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 1 of 1 shared result)
- Spartan: Differentiable Sparsity via Regularized Transportation (NeurIPS 2022 · better on 7 of 7 shared results)
- Powerpropagation: A sparsity inducing weight reparameterisation (NeurIPS 2021 · better on 5 of 5 shared results)
- AC/DC: Alternating Compressed/DeCompressed Training of Deep Neural Networks (NeurIPS 2021 · 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
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.87)
- Sparse maximal update parameterization: A holistic approach to sparse training dynamics (NeurIPS 2024 · similarity 0.86)
- Scaling and evaluating sparse autoencoders (ICLR 2025 · similarity 0.86)
- Navigating Extremes: Dynamic Sparsity in Large Output Spaces (NeurIPS 2024 · similarity 0.86)
- The Journey Matters: Average Parameter Count over Pre-training Unifies Sparse and Dense Scaling Laws (ICLR 2025 · similarity 0.86)
- Pruning Large Language Models with Semi-Structural Adaptive Sparse Training (AAAI 2025 · 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.
