Izawa “Meta-learning Structures a Minimal Meta-Cognitive Architecture for Implicit Motor Adaptation”

Mini-symposium “QB3:Qualia, Brain, Body, Behavior”
2024 January 15 (Mon)
bicr.atr.jp/decnef/mini-sympo...
Jun Izawa
Title: “Meta-learning Structures a Minimal Meta-Cognitive Architecture for Implicit Motor Adaptation”
Abstract:
The motor adaptation process is implicit, occurring without conscious awareness of errors, and it is unavoidable even when it conflicts with instructed task performance. It is considered an automatic process aimed at minimizing sensory prediction errors, that is, reducing the discrepancy between generated and predicted movement trajectories. Therefore, the principle of error minimization is a central premise in computational studies of sensory-motor adaptation.However, it also exhibits great flexibility in the properties of adaptation. For instance, the speeds of motor adaptation are altered by the volatility of environments. Also, it can be influenced by monetary gain and loss. This flexibility has not been explained in a unified manner by the existing theory based on the error minimization principle. Thus, we challenged this conventional view. Our theory of meta-learning, derived from the reward maximization/punishment minimization principle, can promote, and suppress both the speeds of motor adaptation and the retention of motor memories, forming learning-reward associations. It explains the previously reported flexibility in motor adaptation speeds in a unified framework. In addition, we presented empirical evidence of this theory through a novel meta-learning training task for a visuomotor adaptation task. Interestingly, we found both reinforcement learning (i.e., instrumental learning) effects and reinforcement effects (i.e., Pavlovian effects), which go beyond the computational model. This computational model, which achieves ‘learning-to-learn’ in motor learning, parallels a meta-learning architecture recently developed in machine learning science. In this architecture, a higher-level reinforcement learning system trains the lower-level learning, thus structuring the monitoring and control of the lower-level/object-level process by the meta-level process, i.e., the metacognitive system. In this framework, our findings indicate that the human meta-cognitive system can explicitly comprehend implicit motor adaptation, an inherently contradictory interaction between conscious and unconscious cognition.