Research map / Architectures and efficiency / Attention and transformers
Data Distributional Properties Drive Emergent In-Context Learning in Transformers
Stephanie C.Y. Chan, Adam Santoro, Andrew Kyle Lampinen, Jane X Wang, Aaditya K Singh, Pierre Harvey Richemond, James McClelland, Felix Hill
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 |
|---|---|---|---|
| Increase the proportion of bursty training sequences (ablation) | 12-layer transformer · Omniglot holdout classes | in-context learning accuracy | Improved |
| Increase the proportion of bursty training sequences (ablation) | 12-layer transformer · Omniglot trained classes | in-weights learning | Hurt |
| Increase training classes from 100 to 1600 (ablation) | Transformer trained with burstiness fixed at p(bursty)=0.9 · Omniglot holdout classes | in-context learning accuracy | Improved |
| Increase training classes from 100 to 1600 (ablation) | Transformer trained with burstiness fixed at p(bursty)=0.9 · Omniglot trained classes | in-weights learning | Hurt |
| Expand training classes to 12800 using rotated and flipped images (ablation) | Transformer trained with p(bursty)=0.9 · Omniglot holdout classes | in-context learning accuracy | Improved |
| Expand training classes to 12800 using rotated and flipped images (ablation) | Transformer trained with p(bursty)=0.9 · Omniglot trained classes | in-weights learning | Hurt |
| Increase the number of possible labels per image class (ablation) | Transformer trained with p(bursty)=0.9 · Omniglot holdout classes | in-context learning accuracy | Improved |
| Increase within-class image variation (ablation) | Transformer · Omniglot holdout classes | in-context learning accuracy | Improved |
| Increase within-class image variation (ablation) | Transformer · Omniglot trained classes | in-weights learning | Hurt |
| Train on a Zipfian class distribution with exponent 1 (ablation) | Transformer trained with 12800 classes and p(bursty)=0.9 · Omniglot holdout classes | in-context learning accuracy | Improved |
| Train on a Zipfian class distribution with exponent 1 (ablation) | Transformer trained with 12800 classes and p(bursty)=0.9 · Omniglot common classes | in-weights learning | Improved |
| Increase Zipfian skew beyond the uniform distribution (ablation) | Transformer trained with 12800 classes and p(bursty)=0.9 · Omniglot holdout classes | in-context learning accuracy | Hurt |
Closest papers by meaning
- Toward Understanding In-context vs. In-weight Learning (ICLR 2025 · similarity 0.93)
- The mechanistic basis of data dependence and abrupt learning in an in-context classification task (ICLR 2024 · similarity 0.91)
- Breaking through the learning plateaus of in-context learning in Transformer (ICML 2024 · similarity 0.90)
- How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations (ICLR 2024 · similarity 0.90)
- Understanding In-Context Learning in Transformers and LLMs by Learning to Learn Discrete Functions (ICLR 2024 · similarity 0.90)
- In-Context Learning through the Bayesian Prism (ICLR 2024 · similarity 0.90)
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.
