Research map / Label-efficient and robust learning / OOD and anomaly detection
HYBOOD: A Hybrid Generative Model for Out-of-Distribution Detection with Corruption Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HYBOOD, a GLOW-based normalizing flow with a global-average-pooling linear classifier, used to detect covariate and semantic shifts and rank corruption difficulty (CDR) on MNIST-C and CIFAR10-C via Wasserstein distance, mutual information, and MDL.
Problem
Out-of-distribution detection that distinguishes covariate shift (corruptions) from semantic shift, and quantifying how difficult each corruption type is for a generative model.
Approach
A simplified DIGLM-style hybrid flow that replaces the GLM head with global average pooling of latent features plus a linear classifier, combined with a corruption difficulty ranking that sums ranks of Wasserstein distance, mutual information, and minimal description length.
Key results
- HYBOOD achieves F1 of 0.9740 (GAPLM) and 0.9769 (linear) on MNIST, and 0.6778 (GAPLM) and 0.6933 (linear) on CIFAR10 classification.
- Replacing the flatten linear layer with GAP reduces classifier trainable parameters from 15,360 to 240 with small performance loss.
- Corruption difficulty ranking places impulse noise and fog among the most difficult corruptions on MNIST-C and impulse noise and gaussian noise among the hardest on CIFAR10-C.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace GAPLM with a high-level flatten linear layer (ablation) | HYBOOD · MNIST | F1: 0.9740 -> 0.9769 | Helped |
| replace GAPLM with a high-level flatten linear layer (ablation) | HYBOOD · CIFAR10 | F1: 0.6778 -> 0.6933 | Helped |
| replace GAPLM with a high-level flatten linear layer (ablation) | HYBOOD · MNIST | Precision: 0.9751 -> 0.9771 | Helped |
| replace GAPLM with a high-level flatten linear layer (ablation) | HYBOOD · MNIST | Recall: 0.9738 -> 0.9769 | Helped |
| replace GAPLM with a high-level flatten linear layer (ablation) | HYBOOD · CIFAR10 | Precision: 0.6775 -> 0.6931 | Helped |
| replace GAPLM with a high-level flatten linear layer (ablation) | HYBOOD · CIFAR10 | Recall: 0.6782 -> 0.6937 | Helped |
Surprising findings
- Covariate shifts from hard corruptions can lie farther from the in-distribution than the semantic shift dataset in log-density space on MNIST-C.
- The linear classifier slightly outperforms GAP-based head on CIFAR10 (0.6933 vs 0.6778 F1), i.e., downscaling costs little.
Limitations
- Corruption difficulty weights are set heuristically (alpha, beta, gamma = 1) and hyperparameters are heuristically chosen.
- Classification F1 on CIFAR10 is low (~0.68-0.69), limiting the generative-discriminative hybrid's accuracy.
- Evaluation is limited to two ID/OOD dataset groups and no quantitative OOD detection benchmark comparison is reported in the main text.
Experimental setup
Datasets: MNIST MNIST-C Fashion-MNIST CIFAR10 CIFAR10-C SVHN
Benchmarks: MNIST-C CIFAR10-C Fashion-MNIST SVHN
Metrics: F1 precision recall Wasserstein distance mutual information minimum description length log p(x) distribution separability
Baselines: DIGLM
Builds on: DIGLM Glow SCONE
Closest papers by meaning
- Diffusion-based Semantic Outlier Generation via Nuisance Awareness for Out-of-Distribution Detection (AAAI 2025 · similarity 0.86)
- ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.85)
- Rethinking the Evaluation of Out-of-Distribution Detection: A Sorites Paradox (NeurIPS 2024 · similarity 0.85)
- Beyond Clean Training Data: A Versatile and Model-Agnostic Framework for Out-of-Distribution Detection with Contaminated Training Data (CVPR 2025 · similarity 0.84)
- ALTBI: Constructing Improved Outlier Detection Models via Optimization of Inlier-Memorization Effect (AAAI 2025 · similarity 0.84)
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · 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.