tutorial web page
Tutorial in IEEE ICDL-2024
Typical and Atypical Development of Neurorobots Based on Free Energy Principle
Jun Tani, Director of Cognitive Neurorobotics Unit, Okinawa Institute of Science and Technology, https://groups.oist.jp/cnru/jun-tani
Yuichi Yamashita, Section Chief at the Department of Information Medicine, National Center of Neurology and Psychiatry
Abstract
The current tutorial seeks to deepen understanding of how predictive and reflective minds can develop based on the free energy principle (FEP) by taking neurorobotics modeling approach. Especially, the tutorial examines both cases of normal development and abnormal one which is caused by psychiatric disorders such as schizophrenia (SZ) and autism spectrum disorders (ASD). The tutorial begins with an intuitive explanation of FEP, followed by introduction of two essential frameworks derived from FEP, predictive coding (PC) which accounts for perception, and active inference (AiF) for action generation. Thereafter, we explain how these frameworks can be implemented in various types of Bayesian neural networks models with focusing on “typical development of competencies for hierarchical prediction and inference with precision (Bayesian belief) at each level. Next, we introduce a recent approach, so-called computational psychiatry which aims to gain theoretical understanding of developmental disorder mechanisms. We look at computational models for developmental disorders using FEP wherein we elaborate on how malfunctions in hierarchical prediction/inference due to altered belief update could lead to development of SZ and ASD. We introduce a series of neurorobotics studies using the models to showcase some emergent phenomena which have been observed both in typical and atypical development in the embodied experiments.
Contents
The tutorial is provided as a half day lecture course. Contents are as itemized here.
1 Introduction of FEP
1.1 Intuitive explanation of FEP related to cognitive neuroscience and developmental disorder. (Yamashita)
1.2 Mathematical formulation of FEP, PC, and AiF (Tani)
2 Introduction of Bayesian RNN models based on FEP (Tani)
2.1 Basic architecture and mechanisms
2.2 Mathematical and numerical characteristics
3 Typical development of neurorobots (Tani)
3.1 Development of compositionality and hierarchy
3.2 Arbitrating top-down and bottom-up processes by regulating precision structure in goal-directed object manipulation tasks
3.3 The same as above in social interaction tasks
4 Atypical development of neurorobots (Yamashita)
4.1 Altered hierarchical representation and behavioral inflexibility in ASD
4.2 Hierarchy of precisions and traits for SZ and ASD
5 Open discussion with participants
Target audience
The audiences could be both of graduate students and researchers from wide-ranged background including robotics, neural network modeling, neuroscience, machine learning, developmental psychology, and philosophy of minds.
Lecturer
The tutorial lecture will be given by Prof. Jun Tani and Dr. Yuichi Yamashia. JunTani is the director of Cognitive Neurorobotics, Okinawa Institute of Science and Technology in Japan. Jun Tani has studied cognitive neurorobotics using PC and AiF frameworks more than 25 years. He summarized his research results in a book [2]. Yuichi Yamashia is the Section Chief at the Department of Information Medicine, National Center of Neurology and Psychiatry. As a clinical psychiatrist, he has dedicated over 15 years to studying computational psychiatry, employing PC and AiF frameworks.
Acknowledgement
Jun Tani is partially supported by Grant-in-Aid for Transformative Research Area (A)： unified theory of prediction and action（24H02175）.