Smaller AI Models Suffice For Brain-Language Studies, Says IIIT-H Study

Researchers from the International Institute of Information Technology Hyderabad (IIIT-H) found that AI models with about 3 billion parameters perform almost as well as those with up to 14 billion parameters in predicting human brain activity, potentially reducing computing needs for brain-language studies. The study, presented at the International Conference on Machine Learning (ICML) in Seoul, was led by Prof. Bapi Raju and PhD researcher Vijay Rowtula. They tested whether increasing model size improves brain prediction accuracy using techniques like quantisation and pruning. Most techniques reduced model size without substantially affecting brain activity predictions, though performance on some conventional language tasks was affected. The findings challenge earlier research suggesting a 15 per cent improvement in brain prediction accuracy when moving to larger models. Prof. Raju noted a dissociation between brain alignment and linguistic competence, suggesting different abilities are needed for language benchmarks versus brain process modelling. The research could make brain-language studies cheaper and faster due to lower memory and computing requirements of smaller models. At ICML, the researchers exchanged ideas with computational neuroscience researchers from the NeuroAI Lab at EPFL, opening possibilities for future collaboration.