Purpose : Understanding brain dynamics is essential for uncovering how neural systems generate complex cognitive and behavioral functions. Recent advances in neuroimaging, computational modeling, and machine learning have accelerated research into the spatiotemporal organization of brain activity. This workshop will bring together researchers from diverse domains—including systems neuroscience, neuroinformatics, and clinical neuroimaging—to present and discuss current findings and emerging methodologies in brain dynamics research.
Date and time : September 9th (Tuesday), 13:00-17:15
Venue : Online
Session 1 Chair : Yusuke Takeda
13:00-13:05 Opening
13:05-13:35 Resting-State Brain Fluctuations as a Window into Psychiatry: Identifying Generalizable Biomarkers of Psychiatric Disorders, Okito Yamashita (RIKEN/ATR)
13:35-14:05 Dynamic Mode Decomposition for Neural Biomarkers and Decoding, Takufumi Yanagisawa (University of Osaka)
14:05-14:35 The brain characterized as a physical reservoir, Hirokazu Takahashi (University of Tokyo)
14:35-14:50 break
Session2 Chair : Li Yuanhao
14:50-15:15 Nonlinear representation learning for brain-imaging data, Hiroshi Morioka (Shiga University)
15:15-15:40 Simulation-Based Inference and Compensation of Epileptogenic Dynamics in Spiking Microcircuit Models, Daniel Muller-Komorowska (OIST)
15:40-16:05 Brain-JEPA: Brain Dynamics Foundation Model with Gradient Positioning and Spatiotemporal Masking, Zijian Dong (National University of Singapore)
16:05-16:20 break
Session 3 Chair : Hiroshi Morioka
16:20-16:40 Transient interaction of resting-state networks in human spontaneous brain activities, Yusuke Takeda (RIKEN)
16:40-17:00 Non-Gaussian Brain Source Imaging Using Correntropy-Based Improper Noise Model, Yuanhao Li (RIKEN)
17:00-17:05 Closing
---------------- ABSTRACT -------------------------
Resting-State Brain Fluctuations as a Window into Psychiatry: Identifying Generalizable Biomarkers of Psychiatric Disorders, Okito Yamashita (RIKEN Center for advanced intelligence project/ATR)
In this talk, I will present the outcomes of our decade-long collaborative research aimed at developing biomarkers for mental disorders that are generalizable across multiple research centers. Leveraging large-scale resting-state fMRI datasets collected through two major Japanese initiatives—the Strategic Research Program for Brain Sciences (SRPBS, 2012–2018) and Brain/MINDS Beyond (BMB, 2018–2024)—we have successfully developed an objective biomarker for major depressive disorder (MDD) using ensemble sparse classifiers. This biomarker has been rigorously validated with both completely independent datasets and prospective studies. Recent analyses of traveling-subject data have further revealed the computational mechanisms by which the classifier effectively reduces disorder-unrelated variability. In addition, we have developed a stratification biomarker capable of predicting antidepressant treatment response, further advancing the potential for individualized care. Despite these achievements, significant challenges remain in the path toward precision psychiatry. I will conclude by discussing the critical role of computational tools in overcoming these challenges and in shaping the future of biomarker development.
Dynamic Mode Decomposition for Neural Biomarkers and Decoding, Takufumi Yanagisawa (Department of Neuroinformatics, The University of Osaka, Graduate School of Medicine)
We present a novel framework for extracting physiologically meaningful features from neural signals using Dynamic Mode Decomposition (DMD). By decomposing spatiotemporal neural activity into low-dimensional dynamic modes, we obtain two types of interpretable features: spatial (sDM) and temporal-frequency (tfDM) representations. These DMD-derived features outperform conventional spectral features in decoding motor intentions from electrocorticography (ECoG) and in classifying neurodegenerative disorders using EEG. Their ability to capture both spatial and frequency-domain information with high interpretability makes them well suited for real-time brain-computer interfaces and clinical biomarker discovery. Finally, I will introduce our ongoing efforts to integrate optically pumped magnetometers (OPMs) with data-driven digital twin models for non-invasive, personalized brain diagnostics.
The brain characterized as a physical reservoir, Hirokazu Takahashi (Graduate School of Information Science and Technology, The University of Tokyo,
東京大学大学院情報理工学系研究科)
A rich repertoire of neural dynamics is thought to play a crucial role in information processing of complex time-series signals. To quantify the neural dynamics, we sought to use information processing capacity (IPC), an emerging measure in a recent theory of reservoir computation that characterizes the linear and non-linear mapping of past input series within the system. Based on IPC, we discuss whether and how a rich repertoire of intrinsic dynamics in the auditory cortex is related to the origin of music.
Nonlinear representation learning for brain-imaging data, Hiroshi Morioka (Faculty of Data Science, Shiga University)
Brain-imaging data are known to exhibit strong nonlinearity and nonstationarity, which pose significant challenges for conventional analysis methods. To address these issues, we have proposed several frameworks for nonlinear representation learning, particularly in the context of nonlinear independent component analysis, whose assumptions align well with the characteristics of brain signals. In this talk, I will present the key ideas behind our frameworks and demonstrate their effectiveness on real brain-imaging datasets.
Simulation-Based Inference and Compensation of Epileptogenic Dynamics in Spiking Microcircuit Models, Daniel Muller-Komorowska (Neural Coding and Brain Computing Unit, Okinawa Institute of Science and Technology)
Spiking microcircuit models simulate neurons, action potentials, and synaptic transmission to reveal healthy and pathological brain dynamics. These detailed models offer potential for personalized treatments, but identifying parameters that match patient data remains challenging. We apply neural posterior estimation, a modern simulation-based inference method, to efficiently estimate parameter distributions from observed dynamics. Our results show that compensatory mechanisms depend on the underlying epileptogenic cause, suggesting that precision diagnoses combined with simulations could guide optimal treatment strategies.
Brain-JEPA: Brain Dynamics Foundation Model with Gradient Positioning and Spatiotemporal Masking, Zijian Dong (National University of Singapore (Multimodal Neuroimaging in Neuropsychiatric Disorders Laboratory (MNNDL), Centre for Sleep and Cognition (CSC)))
In this talk, I will introduce Brain-JEPA, a brain dynamics foundation model with the Joint-Embedding Predictive Architecture (JEPA). This pioneering model achieves state-of-the-art performance in demographic prediction, disease diagnosis/prognosis, and trait prediction through fine-tuning. Furthermore, it excels in off-the-shelf evaluations (e.g., linear probing) and demonstrates superior generalizability across different ethnic groups, surpassing the previous large model for brain activity significantly. Brain-JEPA incorporates two innovative techniques: Brain Gradient Positioning and Spatiotemporal Masking. These methodologies enhance model performance and advance our understanding of the neural circuits underlying cognition. Overall, Brain-JEPA is paving the way to address pivotal questions of building brain functional coordinate system and masking brain activity at the AI-neuroscience interface, and setting a potentially new paradigm in brain activity analysis through downstream adaptation.
Transient interaction of resting-state networks in human spontaneous brain activities, Yusuke Takeda (RIKEN Center for advanced intelligence project/ATR)
Spontaneous brain activities at rest are organized into resting-state networks (RSNs), such as the default mode network (DMN). While the DMN is assumed to interact with other networks to integrate distributed information, its fine-scale temporal dynamics remain unclear. Here, we demonstrate how the DMN interacts with other networks at millisecond-scale temporal resolution using resting-state magnetoencephalography (MEG) data and our recently proposed method, BigSTeP. As a result, the DMN and the visual/sensorimotor network alternately activated with a lag of about 1/2 π rad in the alpha/beta oscillation. These alternating and orthogonal activities of the RSNs may reflect the brain's strategy to achieve both the integration and separation of visual and sensorimotor information via the DMN.
Non-Gaussian Brain Source Imaging Using Correntropy-Based Improper Noise Model, Li Yuanhao (RIKEN Center for advanced intelligence project)
Brain source imaging using MEG/EEG is highly sensitive to noise and artifacts that often violate the common Gaussian assumption in conventional Bayesian frameworks. In this talk, a novel source imaging approach is introduced which incorporates a correntropy-based improper noise model to enhance robustness against non-Gaussian noise. The method leverages the maximum correntropy criterion (MCC) to define a flexible noise model and applies score matching for hyperparameter estimation without relying on validation data. Combined with hierarchical Bayesian inference, the proposed ChVB algorithm demonstrates significantly improved accuracy in both simulation and real-world datasets, outperforming traditional Gaussian and other heavy-tailed noise models. This work establishes a more reliable foundation for brain source imaging in realistic recording environments.
Public events of RIKEN Center for Advanced Intelligence Project (AIP)
Join community