Human-in-the-loop Concept Discovery and Curation in Vision Foundation Models — H. Lim | HAICON26

Speaker: H. Lim Affiliation: Helmholtz Munich / KAIST AI, South Korea Date: June 9, 2026 Time: 2:34–2:48pm Session: Session 1a – Benchmarking & Testing Recorded at the Helmholtz AI Conference (HAICON26), June 9–11, 2026. More info: https://haicon.cc/conference-program/ About this talk: Human-in-the-loop Concept Discovery and Curation in Vision Foundation Models H. Lim 1 1 Helmholtz Munich / KAIST AI, South Korea This talk introduces PatchSAE (ICLR 2025) for extracting interpretable visual concepts from CLIP via sparse autoencoders, and ConceptScope (NeurIPS 2025) for distinguishing causally important from co-occurring but non-causal concepts. Evaluated on ImageNet, SUN397, and Food101, the approach achieves F1=0.72 for concept detection across six fine-grained datasets.