Research map / Learning theory and algorithms / Kernels and regression theory
Model Similarity Mitigates Test Set Overuse
Horia Mania, John Miller, Ludwig Schmidt, Moritz Hardt, Benjamin Recht
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 measured pairwise model similarity instead of treating model losses as independent | ImageNet ILSVRC models · ImageNet ILSVRC | pairwise loss agreement: 0.62 -> 0.85 | Improved |
| Use a similarity-based refined union bound rather than the standard union bound | ImageNet model test-set reuse · ImageNet validation | testable models: 4 \times improvement | Improved |
| Use pairwise similarity to bound test-set reuse | DARTS random configurations · CIFAR-10 | increase in testable models: 9.9 \times | Improved |
| Assume Naive Bayes structure in addition to using similarity | DARTS random configurations · CIFAR-10 | increase in testable models: 1.6 \cdot 10^4 \times | Improved |
| Use pairwise similarity to bound test-set reuse | DARTS highest-scoring configurations · CIFAR-10 | increase in testable models: 12.0 \times | Improved |
| Assume Naive Bayes structure in addition to using similarity | DARTS highest-scoring configurations · CIFAR-10 | increase in testable models: 3.4 \cdot 10^4 \times | Improved |
| Account for model similarity in the theoretical bound | Non-adaptive evaluation of classifiers · General classification | overfitting guarantee | Improved |
Closest papers by meaning
- Efficient Lifelong Model Evaluation in an Era of Rapid Progress (NeurIPS 2024 · similarity 0.86)
- Reproduce, Replicate, Reevaluate. The Long but Safe Way to Extend Machine Learning Methods (AAAI 2024 · similarity 0.86)
- Benchmark Data Repositories for Better Benchmarking (NeurIPS 2024 · similarity 0.85)
- Does CLIP’s generalization performance mainly stem from high train-test similarity? (ICLR 2024 · similarity 0.85)
- On Local Overfitting and Forgetting in Deep Neural Networks (AAAI 2025 · similarity 0.85)
- What Fits (Into Few Tokens) Doesn't Overfit: Compression and Generalization in ML Research Agents (NeurIPS 2026 · similarity 0.84)
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.
