Training Foundation Models Directly on Human Brain Data (arxiv.org)

Researchers have proposed a novel approach to enhance foundation models by training them directly on human brain data, utilizing neuroimaging to gain insights into cognitive processes that traditional human-generated datasets may overlook. This strategy aims to transcend the limitations of current AI models, which heavily rely on surface-level data, by tapping into deeper neural complexities of human thought. The study identifies four cognitive levels—perception, valuation, execution, and integration—where neuroimaging data could provide valuable information to improve model performance. The authors introduce two innovative methodologies: reinforcement learning from human brain (RLHB) and chain of thought from human brain (CoTHB), which prioritize the strategic use of neuroimaging to enhance foundation model training. This approach not only promises to address existing shortcomings in AI performance but also opens avenues for developing more advanced artificial general intelligence (AGI) systems. Moreover, it raises significant ethical and technical considerations that could impact the future of AI development, creating a potential bridge between current architectures and neuroscience-inspired solutions.