The Language Neuroscience Podcast Encoding and decoding semantic representations with Alexander Huth
May 4, 2023
Alexander Huth, Assistant Professor at the University of Texas, Austin, specializes in fMRI and computational models studying language and meaning. He discusses using natural stimuli for semantic research and how he transitioned from AI to neuroscience. Huth shares insights on creating rich semantic models, mapping semantic responses in the brain, and decoding stories from brain activity. He highlights advancements in neural language models and ethical considerations in brain decoding, revealing how our thoughts can be reconstructed from brain data.
AI Snips
Chapters
Transcript
Episode notes
Manually Labeling Movies Saved The Project
- Alexander Huth manually labeled ~2 hours of movie frames one-second at a time to create reliable semantic labels.
- He preferred doing the laborious work himself instead of messy crowdsourcing to ensure consistent annotations.
Natural Stimuli Shift Effort To Modeling
- Using natural stimuli shifts the scientific effort from experimental design to computational modeling.
- Naturalistic approaches reveal large-scale system behavior that controlled experiments can obscure.
Taxonomy Helps Generalize Semantic Vectors
- Adding WordNet hypernyms makes category vectors generalize and reduces parameter redundancy.
- Shared category relationships act like priors that push related category weights closer together in encoding models.