Abstract:As machine learning becomes increasingly prevalent in impactful decisions, recognizing when inference data is outside the model's expected input distribution is paramount for giving context to predictions. Out-of-distribution (OOD) detection methods have been created for this task. Such methods can be split into representation-based or logit-based methods from whether they respectively utilize the model's embeddings or predictions for OOD detection. In contrast to most papers which solely focus on one such group, we address both. We employ dimensionality reduction on feature embeddings in representation-based methods for both time speedups and improved performance. Additionally, we propose DICE-COL, a modification of the popular logit-based method Directed Sparsification (DICE) that resolves an unnoticed flaw. We demonstrate the effectiveness of our methods on the OpenOODv1.5 benchmark framework, where they significantly improve performance and set state-of-the-art results.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve the detection ability of machine - learning models when facing out - of - distribution (OOD) data. Specifically, when the input data is outside the distribution range seen during model training, how to effectively identify these OOD data to ensure that the model makes predictions only in reliable situations. This is crucial for applications involving critical decisions, such as self - driving cars, medical diagnosis, etc. ### Main contributions of the paper: 1. **Comprehensive improvement of OOD detection methods**: - The author simultaneously explored representation - based and logit - based OOD detection methods, while most existing research usually focuses on only one of them. 2. **Proposing the DICE - COL method**: - Improved the existing logit - based OOD detection method, Directed Sparsiﬁcation (DICE), and solved an unnoticed defect in its design. DICE - COL calculates mask vectors for the weight matrix columns of each category, avoiding the situation where some categories are not predicted at all, thus improving performance. 3. **Introducing dimension reduction techniques**: - Introduced dimension reduction techniques (such as PCA) in the representation - based OOD detection method to reduce the dimension of the feature space. This not only significantly improves the performance of OOD detection but also speeds up the running speed. Dimension reduction makes the distance calculation in high - dimensional space more meaningful, thus improving the effect of OOD detection. ### Experimental results: - The author conducted extensive experiments on the OpenOODv1.5 benchmark framework to verify the effectiveness of the proposed method. In particular, DICE - COL significantly outperforms the original DICE method on multiple datasets, and the dimension - reduced representation methods (such as MDS - PCA, KNN - PCA, RMDS - PCA) also achieve state - of - the - art performance on multiple datasets. ### Summary: This paper significantly improves the model's detection ability when facing OOD data by improving existing OOD detection methods and introducing new technical means (such as dimension reduction). These improvements are of great significance for ensuring the reliability of machine - learning models in practical applications.

Comprehensive OOD Detection Improvements

Look Around and Find Out: OOD Detection with Relative Angles

Recent Advances in OOD Detection: Problems and Approaches

Going Beyond Conventional OOD Detection

OAL: Enhancing OOD Detection Using Latent Diffusion

Logit Scaling for Out-of-Distribution Detection

OpenOOD: Benchmarking Generalized Out-of-Distribution Detection

OpenOOD v1.5: Enhanced Benchmark for Out-of-Distribution Detection

Your Classifier Can Be Secretly a Likelihood-Based OOD Detector

OoDAnalyzer: Interactive Analysis of Out-of-Distribution Samples

General-Purpose Multi-Modal OOD Detection Framework

PnPOOD : Out-Of-Distribution Detection for Text Classification via Plug andPlay Data Augmentation

Out-of-Distribution Detection by Principal Component Correspondence

Classifier-head Informed Feature Masking and Prototype-based Logit Smoothing for Out-of-Distribution Detection

Out-of-Distribution Detection with Overlap Index

Dissecting Out-of-Distribution Detection and Open-Set Recognition: A Critical Analysis of Methods and Benchmarks

ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms

DICE: Detecting In-distribution Contamination in LLM's Fine-tuning Phase for Math Reasoning

Out-of-Distribution Detection with Deep Nearest Neighbors

Pseudo-OOD training for robust language models

TagOOD: A Novel Approach to Out-of-Distribution Detection via Vision-Language Representations and Class Center Learning