What is it about?
Can an AI model that understands images and words already recognize defects, even before it is trained specifically to find them? Our study explores this question through LAKE, a method that uses a small collection of normal images to identify the model’s most useful internal signals. It checks whether parts of a new image differ from normal patterns, then uses descriptions of “normal” and “anomalous” appearances as an additional clue. LAKE produces both an overall anomaly score and a map of where a defect may be, without retraining the model.
Featured Image
Photo by Zoshua Colah on Unsplash
Why is it important?
Finding defects usually requires adapting an AI model or building a large collection of examples. LAKE shows that useful clues may already exist inside a pretrained vision-language model. In tests on three industrial image datasets, it performed strongly at detecting defects and locating them within images. Results on a brain imaging dataset also suggest potential beyond manufacturing. Identifying the internal signals behind its predictions makes the method easier to examine and understand.
Perspectives
What interests me most about this work is the possibility that a pretrained model already contains useful knowledge about abnormalities, even when that knowledge is hard to see in its usual outputs. LAKE offers a way to examine that possibility through a small set of internal signals. I hope this study encourages further research into how reliably we can uncover such capabilities, and where their limits lie.
Chuancheng Shi
University of Sydney
Read the Original
This page is a summary of: Latent Anomaly Knowledge Excavation: Unveiling Sparse Sensitive Neurons in Vision-Language Models, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3767308.3835150.
You can read the full text:
Contributors
The following have contributed to this page
