Can We Ignore Labels In Out of Distribution Detection?

TL;DR

The paper tackles the problem of whether label information can be ignored in out-of-distribution (OOD) detection, framing it through an information-theoretic lens. It proves a Label Blindness Theorem showing that SSL/unsupervised OOD approaches fail when the learning surrogate is independent of in-distribution labels, and introduces the Adjacent OOD benchmark to expose safety gaps where ID and OOD data strongly overlap. Through extensive experiments across supervised, self-supervised, unsupervised, and zero-shot baselines on Adjacent OOD datasets, the authors demonstrate that unlabeled OOD methods often underperform compared to simple baselines, especially when label information is relevant to the OOD task. The work advocates for cautious benchmarking and suggests incorporating limited label information to overcome approximate label blindness, with implications for safety-critical AI systems in real-world settings.

Abstract

Out-of-distribution (OOD) detection methods have recently become more prominent, serving as a core element in safety-critical autonomous systems. One major purpose of OOD detection is to reject invalid inputs that could lead to unpredictable errors and compromise safety. Due to the cost of labeled data, recent works have investigated the feasibility of self-supervised learning (SSL) OOD detection, unlabeled OOD detection, and zero shot OOD detection. In this work, we identify a set of conditions for a theoretical guarantee of failure in unlabeled OOD detection algorithms from an information-theoretic perspective. These conditions are present in all OOD tasks dealing with real-world data: I) we provide theoretical proof of unlabeled OOD detection failure when there exists zero mutual information between the learning objective and the in-distribution labels, a.k.a. 'label blindness', II) we define a new OOD task - Adjacent OOD detection - that tests for label blindness and accounts for a previously ignored safety gap in all OOD detection benchmarks, and III) we perform experiments demonstrating that existing unlabeled OOD methods fail under conditions suggested by our label blindness theory and analyze the implications for future research in unlabeled OOD methods.

Figures (5)

• Figure 1: An example failure case by visualizing the heatmaps of the gradient of a unlabeled SimCLR trained Resnet chen2020simple using the GradCAM method selvaraju2017grad. The OOD detection task is to detect OOD facial expressions. In this case, the OOD detection method fails as justified by our theoretical work, where the representations do not exhibit a strong gradient in regions commonly associated with facial expressions (i.e., eyebrows, mouth, etc.).
• Figure 2: From left to right, sample images from the datasets ICML Facial Expression, Stanford Cars, and Food 101. These datasets contain classes that are visually similar, in contrast to CIFAR10, which includes classes such as airplane and dog that are not visually similar.
• Figure 3: Comparing Images from the CC3M Dataset with captions containing the word angry. These are drastically different from the images in ICML face dataset. Images captioned with other facial expressions also tend to lack a human face.
• Figure 4: Comparing Images from the CC3M Dataset with captions containing the word BMW 3 Series, Dodge Ram, and Honda Odyssey, left to right. These are are quite similar to the images in Cars dataset
• Figure 5: Comparing Images from the CC3M Dataset with captions containing the word Caeser Salad, Donut, and Pizza, left to right. These are are quite similar to the images in Food 101 dataset

Theorems & Definitions (24)

• Definition 2.1
• Definition 2.2
• Theorem 3.1
• Lemma 3.2
• Corollary 3.3
• Theorem 3.4
• Theorem 3.5
• Theorem 4.1
• Proposition C.1
• proof