A self-funded audit of Meta's Tribe v2 model, which predicts fMRI from video and audio, found that it failed to reproduce retinotopic and tonotopic maps. While the model showed significant motion selectivity, the retinotopic results were null unless specific stimulus-geometry controls were applied, suggesting a statistical confound rather than a true positional code. The researcher posits that this "where-vs-what" dissociation might be a generalizable issue in other brain-encoding or video models. AI
IMPACT Highlights potential limitations in current brain-response models and suggests a need for rigorous validation of positional encoding.
RANK_REASON Independent audit of a specific AI model's capabilities and limitations. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
