Research map / Adversarial robustness and security / Certified and empirical robustness
Large Margin Deep Networks for Classification
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 large-margin l2 loss at all layers instead of cross-entropy | 4-hidden-layer CNN · MNIST, 80% label noise | accuracy: 93.9% -> 96.4% | Helped |
| Use large-margin l1 or l2 loss instead of cross-entropy | 4-hidden-layer CNN · MNIST, noisy labels across the tested noise levels | accuracy | Helped |
| Use large-margin l∞ loss instead of cross-entropy | 4-hidden-layer CNN · MNIST, noisy labels | accuracy | Hurt |
| Apply margin at all layers rather than only the input layer (ablation) | MNIST margin models · MNIST noisy labels and other tested tasks | accuracy | Helped |
| Use all-layer l∞ margin instead of cross-entropy | 4-hidden-layer CNN · MNIST, 68 training samples | accuracy: around 3.7% | Helped |
| Use an l∞ margin rather than l1 or l2 margin (ablation) | MNIST margin models · MNIST FGSM and IFGSM attacks | accuracy | Helped |
| Use large-margin l∞ loss rather than cross-entropy (ablation) | 4-hidden-layer CNN · MNIST adversarially trained models, black-box attacks at higher ε | accuracy | Helped |
| Add adversarial training (ablation) | MNIST cross-entropy and margin models · MNIST generalization and noisy-label tasks | accuracy | No clear effect |
| Use large-margin l∞ or l2 loss instead of cross-entropy | ResNet-58 · CIFAR-10 noisy labels across 0%-80% noise | accuracy: 4% to 10% | Helped |
| Use large-margin l1 loss instead of cross-entropy | ResNet-58 · CIFAR-10 with 5% and 1% of training data | accuracy: 2.5% | Helped |
| Use large-margin l1 or l∞ loss instead of cross-entropy | ResNet-58 · CIFAR-10 IFGSM attacks | accuracy: about 14% | Helped |
| Use large-margin loss instead of the prior Lipschitz method | ResNet-58 · CIFAR-10 adversarial examples, SNR 33 | accuracy: 69.1% -> 82% | Helped |
Closest papers by meaning
- Learning Towards The Largest Margins (ICLR 2022 · similarity 0.92)
- Improved Sample Complexities for Deep Neural Networks and Robust Classification via an All-Layer Margin (ICLR 2020 · similarity 0.90)
- Predicting the Generalization Gap in Deep Networks with Margin Distributions (ICLR 2019 · similarity 0.89)
- Reducing Class-Wise Performance Disparity via Margin Regularization (ICLR 2026 · similarity 0.89)
- Input Margins Can Predict Generalization Too (AAAI 2024 · similarity 0.88)
- Detecting Brittle Decisions for Free: Leveraging Margin Consistency in Deep Robust Classifiers (NeurIPS 2024 · similarity 0.86)
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.