AI Briefing[NeurIPS 2023] Out-of-Distribution Detection via Synthetic Outlier Generation (Part 2)
Key point
Introduces recent research that improves OOD detection performance by generating synthetic outliers in the text and image domains.
Details
To overcome the limitation of not being able to see real data for OOD (Out-of-Distribution) detection, this covers research that defines Outlier Generation across various domains.
The first study moves away from the existing approach that relied only on visual outliers, proposing Textual Outlier Exposure, which uses an LLM to generate text-based outliers. It leverages a multimodal network such as CLIP as the backbone, generating textual outliers at three levels (Word, Description, Caption), achieving higher performance at lower cost than visual outliers.
- Word-level: An intuitive design in the form of 'A Photo of {Word}'
- Description-level: Generating class descriptions using GPT-3
- Caption-level: Generating captions containing rich visual semantics via BLIP-2
The second study addresses the Mistaken OOD Generation problem that can occur during the outlier generation process. It focuses on solving the issue where, if the generated outlier has a semantic meaning similar to actual ID (In-Distribution) data, the model confuses the two, resulting in degraded detection performance instead.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.