Research map / Architectures and efficiency / RNNs and normalization
Less Data, Faster Training: repeating smaller datasets speeds up learning via sampling biases
Jingwen Liu, Ezra Edelman, Surbhi Goel, Bingbin Liu
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 |
|---|---|---|---|
| train on a small random subset (2^14) instead of the full population (2^20) with full-batch GD | 2-layer MLP on (20,6) sparse parity · (20,6) sparse parity | steps and compute to converge: 1500 steps vs more than 2000 steps; 100x speedup in compute | Improved |
| remove input bias (zero empirical mean of x, optionally y) from the small dataset (ablation) | small-set training (2^14) on parity · (20,6) parity | accuracy | No clear effect |
| add input bias taken from a small set to online large-set training (ablation) | online training · (20,6) parity | accuracy: speedup only large when fewer than 2^5=32 samples | Mixed |
| use random labels in the first small-set phase before switching to large-set training (ablation) | MLP · parity and SIM | accuracy / norm growth | Improved |
| add per-layer init scaling alpha (first layer divided, second multiplied) (ablation) | MLP full-population training · parity and SIM | small-vs-large gap | Improved |
| use muP parameterization (ablation) | 2-layer width-64 MLP · parity vs SIM | small-vs-large gap | Mixed |
| add separate larger learning rate for W_Q, W_K (ablation) | 2-layer Transformer · (20,6) parity | small-vs-large gap: optimal QK lr 3.6x (large set), 2.5x (100 epochs), 1.25x (6-phase) | Improved |
| increase W_Q, W_K initialization scale (ablation) | 2-layer Transformer · (20,6) parity | small-vs-large gap: optimal 8x default for large set; 2x for 100-epoch; no benefit for 6-phase | Improved |
| add QK normalization (ablation) | Transformer · sparse parity and SIM vs ICL and mod addition | optimization speed / overfitting | Mixed |
Closest papers by meaning
- Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision (NeurIPS 2026 · similarity 0.88)
- Slower Generalization, Faster Memorization: A Sweet Spot in Algorithmic Learning (NeurIPS 2026 · similarity 0.88)
- Compute-Constrained Data Selection (ICLR 2025 · similarity 0.86)
- Repeated Random Sampling for Minimizing the Time-to-Accuracy of Learning (ICLR 2024 · similarity 0.86)
- On the Inductive Bias of Stacking Towards Improving Reasoning (NeurIPS 2024 · similarity 0.86)
- Where is the Truth? The Risk of Getting Confounded in a Continual World (ICML 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.
