The brain is composed of two intertwined aspects: the physical and the information processing. For the physical aspect, large-scale databases of the brain are being constructed by combining the latest measurement and information technologies, such as those used in the Allen Brain Atlas and the Human Connectome Project. In terms of information processing of brain, artificial intelligence inspired by the brain's structure and algorithms has made significant breakthroughs and produced products for general use, like ChatGPT. The concept of the "digital brain" focuses on integrating both of these aspects, databasing the brain and its information processing, to bring forth a new era of neuroscience. Seminars are being organized to further develop this concept.
Organizer: Ken Nakae (ExCELLS), Daisuke Tagami (Kyushu University), Kenji Doya (OIST)
https://www.youtube.com/@kennakae2779/featured
Please enter the following to view the ZOOM LINK of Digital Brain Seminar. We will add your e-mail address to our mailing list at a later date and notify you of the Seminars. Subscribing to the mailing list will be done through Google Groups, so Gmail or similar is preferred. Please note that you may not be able to subscribe to the mailing list if you do not have a Gmail account.
https://docs.google.com/forms/d/1duZBmrP8-1nVFSevK-VwitDGuL3oonJRR4SJCReSnrM/edit
Speaker：Kei Hirose (Kyushu University)
Place: Zoom (Please find the zoom link by the registration)
Date：2025/2/17 10:30〜 (JST)
Title：High-dimensional interpretable factor analysis via penalization
Abstract: Factor analysis is a statistical method for identifying latent factors from the correlation structures of high-dimensional data. It was originally developed for applications in social and behavioral sciences but has since been applied to various research fields, including the natural sciences. An advantage of factor analysis is that it leads to interpretable latent factors, enabling applications such as the identification of active brain regions in neuroscience. In this study, we propose a penalized maximum likelihood estimation method aimed at enhancing the interpretability of latent factors. In particular, the Prenet (Product-based elastic net) penalization allows for the estimation of a perfect simple structure, a desirable characteristic in the factor analysis literature. The usefulness of the proposed method is investigated through real data analyses. Finally, we discuss potential extensions and applications of the proposed method in neuroscience.
Speaker：Se-Bum Paik (Korea Advanced Institute of Science and Technology)
Place: Zoom (Please find the zoom link by the registration)
Date：2025/1/7 15:00-16:30 (JST)
Title：Emergence of Cognitive Functions in Natural and Artificial Neural Networks
Abstract: How do the diverse functions of the brain originate? Understanding the developmental mechanisms that underlie brain functions is a fundamental question in neuroscience, with significant implications for research on artificial neural networks. This talk will introduce principles related to these developmental mechanisms, which differ notably from the data-driven learning paradigms predominantly used in AI. I will present our recent findings demonstrating that early functional circuits and cognitive functions in the brain can emerge spontaneously, even in the complete absence of training. Using a biologically inspired neural network model, I will first show how regularly structured functional maps can arise from simple local interactions between individual cells. I will discuss how evolutionary variations in physical parameters may lead to the development of distinct functional circuitry in the brain. Next, I will demonstrate that higher cognitive functions, such as visual quantity estimation and primitive object detection, can also emerge spontaneously in untrained neural networks. I will argue that random feedforward connections in early circuits may be sufficient to initiate functional circuits. These findings suggest that early visual functions can emerge from the statistical properties of bottom-up projections in hierarchical neural networks, providing insight into the origins of primitive functions in the brain.