RESEARCH

Self-Rewriting Minds

The Computational Neuroscience of Meta-Intelligence

We view humans as hierarchical adaptive systems comprising the brain, body, and environment. The brain does more than simply adapt to changes in the environment. It evaluates its own learning state and selects and updates the memories, skills, and learning strategies that should be used. We refer to this higher-order adaptive capacity as meta-intelligence. By integrating behavioral experiments, computational modeling, neural recording, robotics, artificial intelligence, and causal neural intervention, we seek to uncover its computational principles and neural mechanisms.

Meta-intelligence and meta-learning

01

Meta-Intelligence and Meta-Learning

Learning how to learn

Our capacity to learn is not fixed. The brain evaluates its own learning state and the reliability of its memories, and modifies how it learns in response to reward, punishment, error, and uncertainty. We investigate skill selection, the future retrievability of motor memories, the regulation of learning rates, and the switching between error-based and reward-based learning as computational components of meta-intelligence. In addition to behavioral experiments and computational modeling, we use brain stimulation and causal neural interventions to identify the underlying neural mechanisms.

Reward, decision-making, and motor control

02

Reward, Decision-Making, and Motor Control

The valuation of action

Human movement is not organized solely to maximize accuracy. The brain integrates effort, time, the probability of failure, physical cost, and expected reward when selecting an action. We study motor control as a form of value-based decision-making, focusing on reinforcement learning, motor exploration, reward-prediction errors, and the computation of subjective motor costs. This research contributes to a unified framework of neuromotor economics that connects motor learning with decision-making.

Motor memory and the embodied self

03

Motor Memory and the Embodied Self

How action shapes the sense of self

The sense that we generate our own actions and the sense that our body belongs to us are thought to emerge from the brain’s predictions about bodily movement and their comparison with incoming sensory information. We investigate how motor memories and bodily predictions contribute to the sense of agency, body ownership, and body image. Using virtual reality and robotics, we manipulate the relationship between the body, action, and sensory feedback to uncover the computational principles of the embodied self.

Affective and counterfactual learning

04

Affective and Counterfactual Learning

Learning from what might have been

Humans learn not only from the outcomes they actually receive, but also from counterfactual alternatives—what might have happened had a different action been chosen. Emotions such as regret and relief may serve as learning signals that carry these comparisons into future decisions. We develop computational models of affective decision-making and investigate artificial intelligence systems that learn flexibly from counterfactual information and emotional value.

De novo motor skill learning

05

De Novo Skill Learning

Creating new motor representations

When operating an unfamiliar tool, body, or environment, the brain cannot always rely on modifying an existing motor memory. It may need to construct an entirely new representation of motor commands. We investigate how new coordination patterns and control strategies emerge through exploration of the body’s many degrees of freedom. By examining motor exploration, skill-representation formation, and the reuse or creation of motor memories, we seek to understand how humans acquire genuinely novel skills.

Bayesian pain and body perception

06

Bayesian Pain and Body Perception

Inference, uncertainty, and bodily threat

Pain is not determined solely by the intensity of nociceptive input. The brain integrates prior experience, predictions, visual information, bodily state, and sensory uncertainty to infer the threat currently affecting the body. Using virtual reality, psychophysics, and computational modeling, we investigate how prediction mismatches and Bayesian surprise alter pain perception. We aim to translate these findings into new approaches to pain management that account for individual differences in perception and cognition.