A controversial purported benefit of neural foundation models is to construct representations that are useful for solving AI tasks we cannot yet solve directly [1, 2, 3]. So far, machine learning has progressed largely following the Bitter Lesson [4], solving problems with Big Data and Big Compute. But is it instead ever worth collecting and modeling precious brain data rather than just gathering a bigger data set for our target tasks? When?
Our collaborative team has quite different views on this question: some of us believe strongly in the value of brain representations for AI, even to the point of aiming to found a company on this premise; others of us are strongly skeptical. This proposal aims to formalize and resolve this question together. Compared to year 1, we have doubled the number of PIs working to address the problem. We aim to create a general theory of how learning scales under different tasks and data streams, drawing insights from artificial, simplified models of brains. This will yield theoretical estimates on how well we can learn useful representations from real brain data, and what type of data is best. The intended application, to find smarter algorithms by studying brains, is one of the aspirational pillars of NeuroAI. Because this project directly evaluates a key motivation for, and application of, neural foundation models, this topic is of particular relevance to the Foundation Models working group.
