CAREER: CompCog: Reverse Engineering Human Deliberation with Process-Tracing and Neural Networks

Human cognition involves different modes of thinking, such as planning future actions, recalling past experiences, and reasoning through problems. This project investigates how human brains learn to think efficiently, including choosing which possibilities to consider, which memories to recall, and which reasoning steps to follow. Currently available artificial intelligence systems require enormous amounts of computing power and energy, while the human brain solves similar problems with a tiny fraction of these resources. Understanding how the brain achieves this efficiency provides a blueprint for artificial intelligence that is less costly, more energy-efficient, and better aligned with how people think. The project also helps prepare the next generation of scientists through a new university course, free online training materials that reach thousands of learners, and a program that introduces high school students to how humans and machines think. The project tests the theory that the diverse strategies of human deliberation, including planning, memory retrieval, and reasoning, emerge from a general principle that selects mental actions to maximize reward while minimizing the costs of thinking, such as time and effort. To discover these strategies, the research team trains recurrent neural networks with meta-reinforcement learning to select both mental actions (what to think about) and external actions (what to do). The team then validates these models against rich human data, including choices, response times, and eye movements, which serve as a non-invasive, high-resolution window into ongoing thought. This research aims to: (1) explain how people decide which future scenarios to simulate during planning, and how these strategies adapt to task demands such as time pressure and memory load; (2) explain how people sequentially retrieve and combine memories to make decisions; and (3) explain how flexible reasoning arises when limited neural resources unpack complex computations into sequences of simpler steps over time. Across all aims, the discovered strategies are compared with both human behavior and modern artificial intelligence systems to identify the principles that make human thinking so efficient. All data, code, and trained models are publicly released through an open platform for evaluating computational models of human thinking. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.