GRANTS

This research aims to develop objective diagnostic indicators (biomarkers) for mental disorders by proposing new analysis methods for resting-state fMRI data. To overcome the limitations of conventional functional connectivity analysis between brain regions, it employs deep learning techniques (multilayer perceptron and Transformer models) to achieve high-precision analysis while preserving local spatial pattern information.
Specifically, the research has three main objectives: (1) improving the accuracy of functional connectivity, (2) developing deep learning models for estimating individual characteristics, and (3) creating diagnostic algorithms for mental and neurological disorders. It utilizes data from the Human Connectome Project and the AMED International Brain Project to construct ensemble models that combine conventional and novel methods.
To address the issue of limited sample sizes, a two-stage approach is adopted: creating group models followed by individual optimization. This methodology is applicable to other biological data analyses, and the research team aims to share insights through code sharing and training courses.
Ultimately, the goal is to provide new solutions for precise evaluation of individual brain function characteristics and early diagnosis and treatment of mental and neurological disorders. This research is expected to break through the limitations of conventional analysis methods and bring innovative advancements in the fields of brain science and medicine.

Computational cognitive neuroscience is a new field that combines artificial intelligence technology with functional brain imaging technology, but little research has been conducted on mathematical ability. This project aims to construct a computational model that comprehensively explains the brain information representation of mathematical thinking through a three-step approach: brain data measurement using functional magnetic resonance imaging, feature extraction using artificial neural networks, and integration of brain data and neural networks using encoding models.

This study aims to verify the reduction of physical and mental strain on patients with severe neurological diseases like ALS by using a BMI system to control computer mice, replacing traditional methods like eye-tracking. The technology enables brain-controlled computer operation, facilitating smooth, immediate communication and potentially increasing patients’ social participation and quality of life.

The collective dynamics of complex networks found in matter, life, and society often go beyond what can be explained by individual components and their pairwise relationships, showing emergent behavior governed by higher-order interactions. However, due to the lack of a concise modeling framework for higher-order interactions in complex networks, their roles and functions in collective dynamics remain unclear. This study aims to develop mathematical modeling techniques to describe and analyze complex networks with higher-order interactions, focusing on neural networks on curved statistical manifolds based on Renyi divergence, with the goal of significantly transforming the current modeling framework of complex networks.

While individual consciousness may appear unified, it is actually formed by the highly integrated activity of approximately 100 billion neurons. These neurons constitute various functional modules (such as vision, taste, motor control, etc.), with many systems, including the autonomic nervous system, operating automatically. Recent brain imaging studies have shown that these neural networks are spontaneously active, and through specific conditions (e.g., Alien Hand Syndrome), it’s possible for the left and right hemispheres of the brain to make different decisions. This suggests the potential existence of multiple sensory and decision-making entities within a single individual.
Based on these findings, this research proposes the hypothesis that “consciousness arises from the interaction of parallel operating brain networks” and attempts to verify it using new analytical methods. Specifically, as a natural extension of regression analysis, we will use deep learning models consisting of multiple fully connected layers to model cooperative and alienating interactions between networks, visualize the transition process of consciousness, and elucidate the relationship between network interactions and individual consciousness and behavior.

Surprisal theory quantifies sentence processing load by the probability of occurrence of words relative to context, allowing comparison of large language models and human sentence processing. This project compares the prediction accuracy of surprisal theory with other methods and examines methods that better explain the relationship between large language models and the human brain.

The symbol estrangement hypothesis suggests that the distance between symbols and objects in the brain becomes more distinct as learning progresses. This project examines whether this hypothesis can be simulated on artificial neural networks. We will construct encoding models that predict brain activity using neural network features, and clarify the similarity in the learning process of symbolic systems in the human brain and neural networks.

The haptic processing for vibratory stimulus may partly employ the auditory cortices and vice versa. Our research aims to create a comprehensive map of all the brain regions involved in such processing. A hypothesis where the stimulus can be presented by activation at somatosensory and auditory systems is proposed and investigated through a series of fMRI experiments.

The Maximum Entropy Principle (MEP) is effective for producing unbiased statistical models, yet its standard formulation leaves out many systems of interest. As these setups gain interest, a principled extension of the MEP is necessary. This project uses information-geometry to extend the MEP.