AI BriefingLG AI Research 407
Key point
Introduces research that improves visual OOD detection performance by generating text-based outliers using LLMs.
Details
LG AI Research's DI(Data Intelligence) Lab is conducting research on generating synthetic outliers to improve OOD(Out-of-Distribution) detection performance. In particular, this research moves away from the existing visual outlier exposure approach and proposes Textual Outlier Exposure utilizing LLM(Large Language Model).
The existing approach relied on visual outliers in the image domain, but this had the drawback of significant performance variance and high time consumption when using auxiliary datasets without explicit knowledge of OOD. To address this, the research team built a framework based on multimodal networks such as CLIP that regularizes the detector in the image domain through text.
Textual outliers are generated at three levels:
- Word-level: Intuitive word-based generation in the form 'a photo of {word}'
- Description-level: Detailed descriptions excluding class names, generated using GPT-3
- Caption-level: Captions containing rich visual semantics extracted from visual cues in images via BLIP-2
Experimental results showed that Caption-level textual outliers performed best, because they reflected rich linguistic meaning that incorporated visual elements beyond simple text descriptions. This methodology demonstrated superior performance compared to existing state-of-the-art visual outlier exposure methods on the ImageNet-1K benchmark.
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.