Research map / Learning theory and algorithms / Prediction and decision losses
Learning explanations that are hard to vary
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 the AND-mask instead of averaging gradients | MLP on the synthetic memorization dataset · Synthetic memorization dataset, out-of-distribution test | accuracy | Helped |
| Combine the AND-mask with L1 or L2 regularization | MLP on the synthetic memorization dataset · Synthetic memorization dataset | test accuracy | Helped |
| Use DANN instead of the AND-mask | MLP on the synthetic memorization dataset · Synthetic memorization dataset, out-of-distribution test | accuracy | Hurt |
| Increase the AND-mask agreement threshold (ablation) | ResNet on CIFAR-10 with random labels · CIFAR-10, random-label training set | training accuracy | Hurt |
| Use the AND-mask instead of standard average gradients | ResNet on CIFAR-10 with original labels · CIFAR-10, original labels | test accuracy | No clear effect |
| Use the AND-mask instead of standard average gradients | ResNet on CIFAR-10 with 25% randomly shuffled training labels · CIFAR-10, untouched-label test set | test accuracy | Helped |
| Use the AND-mask instead of regular Adam | ResNet-18 behavioral cloning policy · CoinRun, unseen levels | test loss | Helped |
| Increase the AND-mask threshold (ablation) | Gradient signals on the synthetic memorization dataset · Synthetic dataset, mechanism and shortcut gradient analysis | correlation with mechanism gradients | Helped |
| Use an XOR-mask that removes consistent gradients and keeps sign-disagreeing gradients (ablation) | Gradient signals on the synthetic memorization dataset · Synthetic dataset, mechanism gradient analysis | correlation with mechanism gradients | Hurt |
Closest papers by meaning
- Learning with Explanation Constraints (NeurIPS 2023 · similarity 0.90)
- The Pitfalls of Memorization: When Memorization Hurts Generalization (ICLR 2025 · similarity 0.87)
- Additive Models Explained: A Computational Complexity Approach (NeurIPS 2025 · similarity 0.86)
- Efficient and Accurate Explanation Estimation with Distribution Compression (ICLR 2025 · similarity 0.86)
- To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets (ICLR 2024 · similarity 0.85)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.85)
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.