How Much is Brain Data Worth for Machine Learning?

A new research paper published on arXiv investigates the quantitative value of incorporating neural recordings into machine learning training processes. The study addresses the uncertainty surrounding when and how much brain data benefits model performance compared to traditional task labels. Using a linear Gaussian model, the authors derive mathematical scaling laws that determine performance improvements based on the volume of brain and task samples. They establish exchange rates between neural data and task samples, factoring in variables such as task-brain alignment, noise levels, latent dimensions, and sample sizes. Furthermore, the analysis explores test distribution shifts, identifying conditions where brain-regularized learning enhances robustness through learned invariances. The research ultimately characterizes specific regimes under fixed collection budgets where gathering brain data is economically and technically justified. This theoretical framework provides a foundational understanding for NeuroAI researchers aiming to optimize multimodal estimators, offering clear guidelines on the trade-offs and potential gains of integrating biological neural data into artificial intelligence systems.