Research map / Label-efficient and robust learning / Recognition and long-tail
Constraining Representations Yields Models That Know What They Don't Know
Joao Monteiro, Pau Rodriguez, Pierre-Andre Noel, Issam H. Laradji, David Vazquez
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 |
|---|---|---|---|
| Add TAC to frozen RoBERTa-BASE | RoBERTa-BASE · HWU64, Banking77, and CLINC150 | accuracy (%): 92.6 -> 93.9; 93.9 -> 94.2; 97.0 -> 97.4 | Improved |
| Use TAC (L1) distances to detect prediction errors | RoBERTa-BASE · HWU64 | Det. Rate (%): 81.43 -> 83.25 | Improved |
| Use TAC (L1) distances to detect prediction errors | RoBERTa-BASE · CLINC150 | Det. AUROC (%): 91.36 -> 93.25 | Improved |
| Use TAC (Cosine) scores to detect prediction errors | ViT Base-16x16 · ImageNet | Det. AUROC (%): 80.91 -> 89.41 | Improved |
| Use TAC (L1) distances to flag examples from unseen classes | RoBERTa-BASE · CLINC150 UNKNOWN classes | Det. Rate (%): 91.78 -> 92.82 | Improved |
| Train TAC from scratch rather than use the base classifier | 4-layer CNN · MNIST | Clean accuracy (%): 99.0 -> 98.6 | Hurt |
| Train TAC from scratch rather than use the base classifier | WideResNet-28-10 · CIFAR-10 | Det. AUROC (%): 93.7 -> 95.7 | Improved |
| Use TAC from scratch rather than the base classifier | WideResNet-28-10 · CIFAR-10 | Clean accuracy (%): 95.6 -> 94.9 | Hurt |
| Use a last-layer ECOC ensemble instead of cross-depth TAC (ablation) | RoBERTa-BASE · HWU64 and CLINC150 | accuracy (%): 93.9 -> 73.6; 97.4 -> 83.4 | Hurt |
| Add LCE to the last-layer ECOC approach (ablation) | RoBERTa-BASE · HWU64 and CLINC150 | accuracy (%): 93.9 -> 85.2; 97.4 -> 87.8 | Hurt |
| Use a last-layer ECOC ensemble instead of cross-depth TAC (ablation) | ViT · ImageNet | Det. AUROC (%): 89.4 -> 90.06 | Improved |
| Use a last-layer ECOC ensemble instead of cross-depth TAC (ablation) | ViT · ImageNet | accuracy (%): 80.6 -> 74.4 | Hurt |
Closest papers by meaning
- Initial Guessing Bias: How Untrained Networks Favor Some Classes (ICML 2024 · similarity 0.85)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.84)
- Neural Collapse in Cumulative Link Models for Ordinal Regression: An Analysis with Unconstrained Feature Model (NeurIPS 2025 · similarity 0.83)
- Class Adaptive Conformal Training (NeurIPS 2026 · similarity 0.83)
- Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence (ICLR 2025 · similarity 0.83)
- Error Correction Output Codes for Robust Neural Networks against Weight-errors: A Neural Tangent Kernel Point of View (NeurIPS 2024 · similarity 0.83)
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.
