Exploring the nature of the brain’s computations
Brain-like AI does not think like brains
Modern AI outperforms humans in various tasks, making it increasingly hard to distinguish AI from humans based on their behavior. Such demonstrations have generated stereotypical views: "AI and brains are alike" and "AI vs. brains." This idea is often sold by asking rhetorical questions.
That said, human-like, brain-like AI does not necessarily think like a brain. AI approaches to solving real-world problems differ from those the brains use. The brain would behave differently if it were in AI’s shoes. One way to investigate this issue is to examine fundamental questions that are easy to resolve for the brain but not for AI.
For example, how does the brain/AI
generate a finite set of representations from the infinite experience?
predict the future from current events?
make inferences based on only a few observations?
motivate itself to explore and define new goals?
resolve the stability-flexibility dilemma?
resolve the bias-variance tradeoff?
backpropagate information through time?
Exploring the nature of prefrontal computations
Our research explores the above questions at behavioral, computational, and neural levels. Our lab focuses on developing a theory of how the brain, especially the prefrontal cortex, coordinates multiple brain subsystems, each of which handles problems differently. We refer to this function as prefrontal metacognitive control. Specifically, we examine the computational and neural basis of the following functions:
Reinforcement learning
Metacognitive learning
Predictive cognition
Intuition/inductive bias
Causal inference
Putting together ideas from machine learning and computational neuroscience
We use a proactive approach. The process begins with designing experiments to situate the brain’s computations in the context of AI (Brain↦AI). We then test various brain-inspired models to explain the data collected from our experiments. This step allows us to reverse engineer the brain’s latent computations. In the second step, an independent AI model is trained to create synthetic experiences that generalize original experiments (AI↦Brain). Since this framework broadens human experience at the behavioral and neural level, it will guide us to develop new hypotheses about brain functions, initiating a recursive research process: ((Brain↦AI)↦Brain)↦…. In summary,
Brain↦AI aims to understand how the brain thinks.
AI↦Brain aims to understand why the brain implements such computations.
((Brain↦AI)↦Brain)↦… creates a deeper understanding of the nature of the brain’s computations.
Reverse and forward engineering human intelligence through the lens of AI
We hope that our investigation will
build machine learning models with brain-like traits (neuroscience-inspired AI*),
solve psychiatric disorders (computational psychiatry), and ultimately
elucidate the nature of human intelligence (computational neuroscience),
Acknowledgments
Our research is supported by the National Research Foundation of Korea and the Ministry of Science and ICT - SW StarLab (2023-2030). We also have received generous support from Google DeepMind (2017), IBM Research (2021), Microsoft Research (2023), Samsung (2016-2023), LGE (2023-2025), KT (2021-2024), etc.
*refer to the summary slides. For more details, please visit our research center.