[논문 리뷰] Why Normalizing Flows Fail to Detect Out-of-Distribution Data

Polina Kirichenko, Pavel Izmailov외 1명|arXiv (Cornell University)|2020. 06. 15.

Generative Adversarial Networks and Image Synthesis참고 문헌 6인용 수 31

한 줄 요약

이 논문은 정규화 흐름(normalizing flows)이 왜 종종 OOD 데이터를 분포 외 데이터로 오분류하는지 분석하고, 그들의 inductive biases가 시맨틱 구조가 아닌 지역 픽셀 상관관계에 의존하게 만든다는 것을 보여준다; 또한 OOD 탐지를 개선하기 위한 아키텍처 변화와 시맨틱 특징의 이점을 강조한다.

ABSTRACT

Detecting out-of-distribution (OOD) data is crucial for robust machine learning systems. Normalizing flows are flexible deep generative models that often surprisingly fail to distinguish between in- and out-of-distribution data: a flow trained on pictures of clothing assigns higher likelihood to handwritten digits. We investigate why normalizing flows perform poorly for OOD detection. We demonstrate that flows learn local pixel correlations and generic image-to-latent-space transformations which are not specific to the target image dataset. We show that by modifying the architecture of flow coupling layers we can bias the flow towards learning the semantic structure of the target data, improving OOD detection. Our investigation reveals that properties that enable flows to generate high-fidelity images can have a detrimental effect on OOD detection.

연구 동기 및 목표

• Identify why normalizing flows fail to detect OOD data despite exact likelihoods.
• Analyze the inductive biases of coupling-layer flows that hamper OOD detection.
• Visualize latent representations to understand what semantic information flows capture.
• Propose architectural adjustments to coupling layers and masking strategies to bias flows toward semantic structure.
• Evaluate the impact of semantically rich embeddings on OOD detection with flows.

제안 방법

• Analyze log-likelihood behavior of RealNVP/flow-based models on in- and out-of-distribution datasets (e.g., ImageNet, CelebA, SVHN).
• Visualize intermediate coupling layer activations and learned s (scale) and t (shift) parameters to understand local vs. semantic representations.
• Investigate how coupling-layer mechanisms (local pixel correlations and co-adaptation) drive high likelihoods for both in- and out-of-distribution data.
• Experiment with masking strategies (checkerboard, horizontal, cycle-mask) to assess effects on OOD detection.
• Introduce bottlenecks in the st-networks to limit capacity and reduce dependence on local pixel correlations.
• Demonstrate improved OOD detection when flows are trained on high-level semantic embeddings rather than raw pixels.

실험 결과

연구 질문

• RQ1 What inductive biases in normalizing flows cause poor OOD detection?
• RQ2 Do local pixel correlations and coupling-layer co-adaptation enable flows to assign high likelihood to OOD data?
• RQ3 Can architectural changes to coupling layers or masking strategies shift flows toward semantic representations to improve OOD detection?
• RQ4 Does training on high-level semantic embeddings enhance OOD detection compared to raw pixel training?

주요 결과

• Flows learn latent representations based on local graphical structure rather than semantic content, hindering semantic OOD detection.
• Coupling layers can predict masked pixels accurately on both in- and out-of-distribution data, driving high likelihood for diverse structured images.
• Changing masking strategies or introducing bottlenecks in st-networks reduces coupling-layer co-adaptation and improves OOD ranking of in-distribution data.
• Training flows on image embeddings with semantic content markedly improves OOD detection compared to raw-pixel training.
• OOD detection improves when using high-level features (embeddings) rather than raw pixel data, achieving clearer separation between in-distribution and OOD datasets.