Foundation models, that are pre-trained by massive data then serve as the core of various systems fine-tuned to accomplish a range of specific tasks, have begun to be deployed in Neuroscience. Here I consider how this new approach may go beyond input-output predictions to yield mechanistic understanding, and enable superb mental abilities which are still poorly understood and elusive in artificial intelligence (AI) systems. I propose the development of a computational platform that incorporates connectome and cell types, captures complex dynamics of recurrent neural circuits, and learns to accomplish multiple cognitive tasks the combined performance of which measures intelligence. Illustrated by recent works in the field of NeuroAI, key requirements are outlined for a biologically-based foundational brain model of intelligence with, at its core, the prefrontal cortex that underlies cognitive capabilities ranging from learning-to-learn, reasoning, planning to executive control of behavior. Resulting novel algorithms could inspire the advancement of general human-like AI.
Xiao‐Jing Wang (Thu,) studied this question.