Research map / Label-efficient and robust learning / Recognition and long-tail
Do Input Gradients Highlight Discriminative Features?
Harshay Shah, Prateek Jain, Praneeth Netrapalli
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 |
|---|---|---|---|
| train with l2/l_inf adversarial robustness instead of standard training (ablation) | MLPs and Resnets · SVHN, Fashion MNIST, CIFAR-10, ImageNet-10 | DiffROAR: -5 -> 15 | Improved |
| adversarially robust MLPs vs standard MLPs at k=10% (ablation) | MLP · SVHN | DiffROAR: roughly 10-30% | Improved |
| standard Resnets on CIFAR-10 and ImageNet-10 at k<40% | standard Resnet · CIFAR-10, ImageNet-10 | DiffROAR | Hurt |
| adversarial training of SVHN MLP at k=5% (top-k vs bottom-k predictive power) (ablation) | MLP, l_inf eps=2/255 · SVHN | predictive power: roughly 40% | Improved |
| adversarial robust model vs standard model on BlockMNIST attributions (ablation) | MLP and Resnet18 · BlockMNIST | attribution on signal vs null block | Improved |
| lower learning rate 0.1 to 0.01, weight decay 0.0001 to 0, momentum 0.9 to 0 (ablation) | retraining on masked data · CIFAR-10 | DiffROAR: 0.1 -> 0.01 | No clear effect |
| increase PGD adversarial training perturbation budget epsilon (ablation) | adversarially trained models · CIFAR-10 | DiffROAR | Improved |
| evaluate BlockMNIST-Top-trained models on BlockMNIST-Bottom | Resnet · BlockMNIST | test accuracy (%): 100 -> 55 | Hurt |
| adversarial training increases robustness but lowers clean accuracy | MLPs and CNNs · SVHN, Fashion MNIST, CIFAR-10, ImageNet-10 | standard test accuracy | Hurt |
Closest papers by meaning
- Discriminative Score Function: Turning Pretrained Models into Functional Generative Priors (NeurIPS 2026 · similarity 0.85)
- Not All Diffusion Model Activations Have Been Evaluated as Discriminative Features (NeurIPS 2024 · similarity 0.85)
- On the Foundations of Shortcut Learning (ICLR 2024 · similarity 0.85)
- Unlearning-based Neural Interpretations (ICLR 2025 · similarity 0.84)
- From Black-box to Causal-box: Towards Building More Interpretable Models (NeurIPS 2025 · similarity 0.84)
- Beyond Interpretability: The Gains of Feature Monosemanticity on Model Robustness (ICLR 2025 · 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.
