Episode · Dec 9, 2025
Why Vision Language Models Ignore What They See
The TWIML AI Podcast
We use cookies for analytics.
Podcast, The TWIML AI Podcast
Hosted by Sam Charrington
Jason Corso discusses how to evaluate zero-shot labels for computer-vision datasets. He describes experiments comparing labels produced by foundation models with human annotations, ways to prioritize examples for review, and what happens when a new class falls outside the training distribution.