Research map / Label-efficient and robust learning / Semi-supervised learning
Max-Margin Deep Generative Models for (Semi-)Supervised Learning
Chongxuan Li, Jun Zhu, Bo Zhang
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 |
|---|---|---|---|
| add max-margin hat loss regularization on unlabeled data (ablation) | Conv-MMCVA · MNIST (100 labels) | error rate (%): 6.4 -> 1.24 | Improved |
| add label-balance regularization (ablation) | Conv-MMCVA · MNIST (100 labels) | error rate (%): reduce 0.2% | Improved |
| replace Pegasos hinge classifier on VA with max-margin training jointly with the generative model (CNN) | Conv-VA · MNIST (full labels) | error rate (%): 1.35 -> 0.45 | Improved |
| replace Pegasos with max-margin joint learning (CNN) | Conv-VA · SVHN (full labels) | error rate (%): 25.3 -> 3.09 | Improved |
| replace Pegasos with max-margin joint learning (MLP) | VA · MNIST (full labels) | error rate (%): 1.04 -> 0.90 | Improved |
| raise margin weight C from 0 to 10^3 (ablation) | Conv-MMVA · MNIST (full labels) | error rate (%): 1.35 -> 0.45 | Improved |
| set C=1 instead of 0 (ablation) | Conv-MMVA · MNIST (full labels) | error rate (%): 1.35 -> 1.86 | Hurt |
| increase C from 0 to 10^4 (ablation) | Conv-MMVA · MNIST (full labels) | variational lower bound: -93.17 -> -112.12 | Hurt |
| increase C from 10^3 to 10^4 (ablation) | Conv-MMVA · MNIST (full labels) | error rate (%): 0.45 -> 0.43 | Improved |
| use max-margin DGM with inferred missing values instead of CNN on incomplete input | Conv-MMVA · MNIST Rect 12x12 missing | error rate (%): 47.2 -> 15.9 | Improved |
| add max-margin to Conv-VA for imputation under rectangle missing | Conv-VA · MNIST Rect 12x12 missing (imputation) | MSE: 0.1299 -> 0.1090 | Improved |
| add max-margin to Conv-VA for imputation under random pixel drop 0.2 | Conv-VA · MNIST Rand-Drop 0.2 (imputation) | MSE: 0.0111 -> 0.0147 | Hurt |
Closest papers by meaning
- Recent Advances in Large Margin Learning (TPAMI 2021 · similarity 0.88)
- Deep Generative Modeling on Limited Data with Regularization by Nontransferable Pre-trained Models (ICLR 2023 · similarity 0.88)
- Classification Accuracy Score for Conditional Generative Models (NeurIPS 2019 · similarity 0.88)
- Generative Marginalization Models (ICML 2024 · similarity 0.87)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.86)
- Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (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.
