Research map / Learning theory and algorithms / Prediction and decision losses
Lower Bounds on Cross-Entropy Loss in the Presence of Test-time Adversaries
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 |
|---|---|---|---|
| Train with clipped optimal soft labels instead of hard labels | ResNet-18 · FMNIST ε=4.6 train | robust cross-entropy loss: 0.349 -> 0.326 | Helped |
| Train with clipped optimal soft labels instead of hard labels | ResNet-18 · FMNIST ε=4.6 test | robust cross-entropy loss: 0.348 -> 0.331 | Helped |
| Train with clipped optimal soft labels instead of hard labels | ResNet-18 · FMNIST ε=5.0 train | robust cross-entropy loss: 0.451 -> 0.419 | Helped |
| Train with clipped optimal soft labels instead of hard labels | ResNet-18 · FMNIST ε=5.0 test | robust cross-entropy loss: 0.451 -> 0.420 | Helped |
| Replace ReLU with ELU (ablation) | ResNet-18 · MNIST ε=3.0 | robust cross-entropy loss: 0.106 -> 1.056 train; 0.236 -> 1.060 test | Hurt |
| Replace ReLU with Tanh (ablation) | ResNet-18 · MNIST ε=3.0 | robust cross-entropy loss: 0.106 -> 13.012 train; 0.236 -> 13.099 test | Hurt |
| Replace ReLU with Leaky ReLU (ablation) | ResNet-18 · MNIST ε=3.0 | robust cross-entropy loss: 0.106 -> 0.103 train; 0.236 -> 0.348 test | Mixed |
| Replace ReLU with SELU (ablation) | ResNet-18 · MNIST ε=3.0 | robust cross-entropy loss: 0.106 -> 0.704 train; 0.236 -> 0.706 test | Hurt |
| Use ResNet-50 instead of ResNet-18 (ablation) | Robust training · FMNIST ε=5.0 | robust cross-entropy loss: 0.451 -> 0.387 train; 0.451 -> 0.387 test | Helped |
| Use ResNet-101 instead of ResNet-18 (ablation) | Robust training · FMNIST ε=5.0 | robust cross-entropy loss: 0.451 -> 0.422 train; 0.451 -> 0.425 test | Mixed |
| Use ResNet-101 instead of ResNet-18 (ablation) | Robust training · FMNIST ε=5.0 | robust cross-entropy loss: 0.42 compared with 0.45 | Hurt |
| Use the custom recursive algorithm rather than a general-purpose convex solver | Lower-bound computation · CIFAR-10 | speedup: up to 3000 × | Helped |
Closest papers by meaning
- Characterizing the Optimal $0-1$ Loss for Multi-class Classification with a Test-time Attacker (NeurIPS 2023 · similarity 0.88)
- Cross-Entropy Loss Functions: Theoretical Analysis and Applications (ICML 2023 · similarity 0.88)
- Robust Loss Functions for Training Decision Trees with Noisy Labels (AAAI 2024 · similarity 0.86)
- A Universal Growth Rate for Learning with Smooth Surrogate Losses (NeurIPS 2024 · similarity 0.86)
- Generative Cross-Entropy: A Strictly Proper Loss for Data-Efficient Classification (NeurIPS 2026 · similarity 0.86)
- Exactly Tight Information-theoretic Generalization Bounds via Binary Jensen-Shannon Divergence (ICML 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.