Neural Computation Workshop 2026 (FY2025)

Aim:

The aim is for current and former members of Doya unit to exchange recent progress and new ideas.

Timeline:

Travel expense support application deadline: Closed
Application as a speaker deadline: Closed
Abstract submission for speaker's deadline: March 17, 2026
On-site registration deadline extended: Closed

Online registration:　To participate online, you must register via the Zoom link below

Program

9:00 Registration
9:20 Opening

Session 1 (chair: Yukako Yamane)
 9:30 Ekaterina Sangati
10:00 Wataru Toyokawa (online)
10:30 Yuji Kanagawa
10:50 break

Session 2 (chair: Jovan Robolledo-Mendez)
11:10 Eiji Uchibe
11:40 Viktor Zhumatiy
12:10 Kevin Max

12:30 Lunch & Poster Session

Poster presenation List
#1 Razvan Gamanut
      "A computational bottom-up, mesoscale approach for the study of the claustrum dynamics
#2 Bogna Ignatowska-Jankowska
      "Disruption of kinematic phenotypes induced by cannabinoid agonist and harmaline in marker-based 3D motion capture of freely moving mice"
#3 Jianning Chen
      "Fictive Learning in Model-based Reinforcement Learning by Generalized Reward Prediction Errors"
#4 Vitória Yumi Uetuki Nicoleti
      "Dynamic Causal Modeling of Marmoset Calcium Imaging Data During Learning of a Reaching Task"
#5 Ayan Chakraborty
      "Probabilistic Inference in Spiking Neural Networks."
#6 Tojoarisoa Rakotoaritina
      "A unified information theoretic framework for intrinsic rewards in reinforcement learning."
#7 Shuhei Hara
      "Top-down Sharpening in Schizophrenia: An fMRI study with Deep Neural Network-Based Decoding"
#8 Yukako Yamane
      "Diverse cortical dynamics underlying body motion in feely moving marmosets"
#9 Katsuhiko Miyazaki
      "Serotonin neurons in the dorsal raphe nucleus encode probability but not amount of future rewards during operant waiting for delayed rewards."
#10 Inaya Rahmanisa
       "Training RNN with Teacher Forcing Strategies for Reconstructing Calcium Imaging Dynamics"
#11 Hideyuki Yoshimura
       "Hierarchical Control under Asymmetric State Information Access"
#12 Yusaku Kasai
       "Toward a Unified Cortical Model Based on Predictive Coding: Integrating Action Intention, Context, Sensory Inputs, and Motor Control"
#13 Jovan Rebolledo Mendez
      "Predicting Parkinson's Disease Progression from a Single Baseline PET Scan Using Enhanced Network Diffusion Models in Marmosets"
#14 Yui Tateyama
      "Analysis on Multi-Site and Multi-Disorder fMRI Data With Brain Foundation Model"

Session 3 (chair: Katsuhiko Miyazaki)
14:00 Akihiro Funamizu
14:30 Masakazu Taira (online)
15:00 Jianning Chen
15:20 break

Session 4 (chair: Ayan Chakraborty)
15:40 Junichiro Yoshimoto
16:10 Alan Fermin
16:40 Bogna Ignatowska-Jankowska
17:00 break

Session 5 (chair: Jeanne Barthelemy)
17:20 Jun Igarashi
17:50 Carlos Gutierrez
18:20 Razvan Gamanut
18:40 Closing
19:00 Dinner

------------------------

Abstract:
Watru Toyokakawa
RIKEN

Connecting reinforcement learning models with social learning and collective behaviour

Collective behaviour consists of multiple dynamical processes where individuals update their beliefs and actions through the repetitions of experience and social interactions. The updated knowledge will then drive new collective behavioural patterns, shaping the form of social interactions. Such a "time-depth" perspective to collective decision making or cultural evolutionary dynamics fits well with the reinforcement learning framework. In this talk, I would like to present some basic overviews of RL models used in human social learning research, and show that RL can provide useful building blocks in understanding collective decision dynamics. I will then present my previous work showing that learning agents, who would have been prone to suboptimal risk aversion, can be “rescued” by the “copy-majority” strategy that promotes more exploration in the group. I would like to discuss some possible ways to extend such RL-based social learning models to more complex group situations.

Akihiro Funamizu
Institute for Quantitative Biosciences, the University of Tokyo

How does experience shape choice behavior in mice?

Our decision-making relies on prior knowledge, especially when sensory inputs are uncertain. Optimal decisions that integrate state estimation and sensory inputs are often characterized within a Bayesian inference framework. Our lab now focuses on how prior experience alters choice behavior in mice. We use head-fixed mice and are developing three behavioral tasks to test the effects of prior experience on decision-making. In some cases, we use artificial neural networks (ANNs) to model mouse choice behavior. In parallel, we are developing a wide-field two-photon microscope (diesel2p) as part of a collaborative study for neural imaging. By integrating mouse behavioral tasks, ANN-based behavioral modeling, and neural imaging, we aim to integrate how prior experience shapes choice behavior in mice.

Masakazu Taira
The University of Sydney

Investigating the neural circuits of model-based behaviours as differentially dictated by cues that are distal or proximal to rewards

Through learning, we develop cognitive maps of our environment that comprise associations between sensory stimuli and rewards. This enables us to predict rewards based on complex features of our environment, referred to as model-based behaviour. Model-based behaviours are important for flexibly adapting choices in changing environments. Some sensory cues directly predict rewards (proximal cues), while others predict reward indirectly through associations with proximal cues (distal cues). Our recent studies demonstrate that the lateral hypothalamus (LH) biases model-based learning and behaviour towards proximal cues and explicitly away from distal cues. In contrast, based on accumulating evidence, we hypothesised that the lateral orbitofrontal cortex (lOFC) is important for inferring model-based paths to reward, regardless of their proximity to reward. To test and compare the role of LH and lOFC in navigating through cognitive maps, we adapted the Daw two-step task, which is used to test the humans’ ability to employ model-based behaviours. Rats first receive one of two distal cues that are further from reward followed by presentation of two levers. Rats press one of the levers and then receive one of two proximal cues. The distal cues inform different state transitions from the levers to the proximal cues. In turn, the proximal cues inform the fluctuating reward probabilities (high/low). Once the rats are appropriately choosing the lever with the higher chance of rewards, they experience one of two reversals between: 1) the reward probabilities of the proximal cues, or which distal cue is being presented in a block. Therefore, rats need to keep responding flexibly and locate the correct proximal cue by making appropriate actions given the presented distal cues. We found that rats can appropriately use the transitional structure of the task to guide their choices. Critically, our two-step task allows us to selectively inhibit distinct neuronal populations in our model-based task during either proximal or distal components of cognitive maps, which differentiates our task from prior rodent versions. Accordingly, we examined the effect of optogenetically inhibiting lOFC during either distal cues, proximal cues, or outcome receipt. We found that inhibition of lOFC during distal cues or outcome delivery disrupted model-based behaviour. However, inhibition of lOFC during the proximal cue had no effect. This is surprising because the proximal cue point is when rats must interpret the meaning of the trial and requires inference of past states to predict the future (i.e., it is a non-Markovian state), which is thought to be dependent on lOFC. Instead, these data may indicate that lOFC is important when behaviour needs to reflect integration of knowledge about the task (i.e., decision making).  We are now performing computational modelling with tiny recurrent neural network (RNN) models to reveal how lOFC is contributing to behaviour in task. Further, we are currently running comparable optogenetic studies with LH to compare the function of these regions.

Junichiro Yoshimoto
Fujita Health University

Data-Driven Stratification of Depression Based on Resting-State Functional Connectivity: From Methodological Foundations to Recent Advances

Major depressive disorder (MDD) is a highly heterogeneous psychiatric condition characterized by diverse symptom profiles, clinical courses, and treatment responses. Currently, practical diagnostic frameworks rely on symptom-based criteria, which do not necessarily reflect the underlying neurobiological mechanisms. This limitation has motivated efforts to identify biologically meaningful subtypes of depression using neuroimaging and data-driven approaches. Resting-state functional connectivity (rs-FC) derived from functional magnetic resonance imaging (fMRI) has emerged as a promising source of biomarkers that capture large-scale brain network organization. This presentation reviews our challenge and recent progress in data-driven stratification of depression based on rs-FC.

We first introduce an early methodological framework that integrates neuroimaging features with clinical and behavioral data to uncover latent subtypes of depression. Using a multiple co-clustering approach applied to high-dimensional datasets consisting of rs-FC patterns, clinical questionnaire scores, and biological measures, three neurophysiological subtypes of depression were identified. These subtypes were characterized by distinct connectivity patterns involving the angular gyrus within the default mode network and were associated with childhood trauma scores. Importantly, these subtypes also showed differences in treatment response to selective serotonin reuptake inhibitors (SSRIs), suggesting that data-driven neurobiological stratification may provide clinically relevant information beyond traditional symptom-based diagnoses.

Despite these promising findings, a major challenge in biomarker research has been the lack of reproducibility and generalizability across independent datasets. Measurement variability in fMRI data and overfitting to discovery cohorts often limit the applicability of proposed biomarkers. To address this issue, more recent work has focused on developing generalizable stratification biomarkers. In this framework, diagnostic biomarkers are first constructed using supervised learning, and subsets of highly informative functional connections are then used to perform unsupervised stratification of patients.

Using this hierarchical strategy, a robust stratification biomarker based on thalamo–somatomotor functional connectivity was identified. Patients stratified by this biomarker exhibited significantly different responses to antidepressant treatment, indicating that specific thalamocortical connectivity patterns may reflect biologically meaningful subtypes of depression relevant to treatment selection. Importantly, the biomarker demonstrated reproducibility across independent cohorts, highlighting the importance of methodological designs that explicitly address generalizability. We expected those achievements would contribute to precision psychiatry, where neurobiological markers derived from brain connectivity can inform individualized diagnosis and treatment strategies.

Ekaterina Sangati
RIKEN

Learning from Collective Feedback

Adaptive systems—from multicellular organisms to human groups—function through the coordinated actions of many interacting components, often without central control. Yet it remains poorly understood how such coordination arises when goals and performance feedback are available only at the collective level. How can groups learn and adapt under such conditions?

In this talk, I present ongoing work investigating collective learning under uncertainty using a dynamic multi-agent multi-armed bandit (MAMAB) task with global rewards. Combining agent-based simulations with online experiments, we examine how simple learning and social influence mechanisms enable groups to coordinate. Our results suggest that positive feedback between individual learning and social influence can support collective adaptation, but that performance depends strongly on group size. Human groups show distinctive dynamics that diverge from model predictions, suggesting that different cognitive and social mechanisms may operate in different interactive contexts. This raises conceptual challenges for scaling theories of social cognition—often developed in dyadic contexts—to larger groups, and highlights the need for tighter integration between modeling and experiments.

Yuji Kanagawa
OIST

Evolution of Fear and Social Rewards in Prey-Predator Relationship

Fear is a critical brain function that enables us to learn to avoid danger via reinforcement learning (RL). However, how fear has evolved under varying predatory pressures remains an open question. In this study, we investigate the relationship between predatory pressure and fear using an evolutionary simulation of RL agents with evolving rewards. In our simulation, prey and predator RL agents co-evolve their reward functions, including visual rewards for observing prey and predators. While fear-like negative rewards for observing predators often evolved in prey, positive rewards for both predators and prey enhancing the grouping behavior sometimes evolved, especially when the predators are not very active. A comparison between different environmental conditions revealed that stronger predator hunting capability promoted stronger fear reward, while less food supply promoted more negative social reward. Moreover, fear did not evolve in response to static pitfalls with non-lethal damage, suggesting that actively hunting predators played an important role in its evolution. These results highlight the special role of predators in the diverse evolution of fear and social rewards.

Jianning Chen
OIST

Meta-learning in biological intelligence: Neurocomputational Mechanism of Behavioral Strategy Regulation.

Regulating behavioral strategies for adaptive learning and decision-making in response to changing external environments and internal states is central to meta-learning, the capability of learning to learn. Multiple behavioral strategies are observed in humans and animals, including greedy reward optimization, novelty seeking, choice preservation, and model-based decision-making, as well as the patterns of change among them. Behavioral strategy regulation might be implemented as either continuous strategy change, implemented as changes in the meta-parameters, the parameters regulating learning processes, or discrete strategy switching. How such regulatory mechanisms are implemented, how they are affected by external and internal factors, and how neuromodulatory systems such as serotonin contribute to them remain unclear.

Using computational modeling, simulation, and a rodent sequential decision-making task, this study investigated the neurocomputational mechanism of behavioral strategy regulation with a focus on the role of serotonin. Dynamic reinforcement learning modeling and state-space modeling show that continuous drift and discrete switches operate on different timescales to jointly shape strategy regulation. It also identified the factors that modulate strategy regulation, highlighting how changes in internal state in response to environmental changes affect strategy selection. Lastly, optogenetic inhibition of serotonin neurons in the dorsal raphe nucleus revealed that serotonergic inhibition suppresses model-based counterfactual learning.

Razvan Gamanut
OIST

Claustrum, retrosplenial cortex and their interaction with serotonin during the delayed reward task in mice

The claustrum is a thin structure densely connected with the brain. Notably, with the default mode network (DMN) it has both a strong anatomical connectivity and, during resting states, a strong functional connectivity.
Here, we are observing the behaviour of the claustrum neurons and the activity of the retrosplenial cortex (RSP, an important component of the DMN) during waiting for a long period for reward. RSP is involved in either spatial-oriented functions by translating between self-perspective and allocentric perspective, or non-spatial functions by comparing perceptual input with memory. Because the claustrum targets non-spatial processing cells in RSP, we hypothesized that the behaviour of the two would be correlated during waiting for reward, when the demand for spatial information is minimum.
Serotonin was also found to increase the functional connectivity of the DMN. In this project we are testing primarily how the serotonin interacts with the neurons in the claustrum that project to retrosplenial cortex, during the delayed reward task.

Kevin Max
OIST

Few-shot, continual learning for spiking neuromorphic olfaction

Neuromorphic olfaction combines sensing of chemical signals with brain-inspired circuit
architectures to emulate key computational principles of biological olfactory systems. This
approach holds strong promises for real-life applications, including detection of dangerous
compounds, air-quality monitoring, and health diagnostics. However, real-world
deployment remains constrained by critical limitations: lack of robust few-shot learning
and class-incremental continual learning algorithms, particularly under the constraints set
by the sensing and processing hardware. Here, we introduce Spi-Fly, a spiking neural
network architecture inspired by the olfactory circuit of Drosophila. Spi-Fly combines
high-dimensional sparse coding with an associative memory mechanism, enabling rapid
few-shot learning, stable class-incremental continual learning without backpropagation, and
operates effectively under low-bit precision. Our results suggest that fruit fly-inspired
sparse associative learning provides a hardware-ready pathway toward fast, continual, and
energy-efficient neuromorphic olfactory intelligence.

Bogna M Ignatowska-Jankowska
OIST

Task-dependent neuronal activity in cerebellar nuclei of freely moving mice performing complex behaviors in 3D

In my previous studies, I used a marker-based 3D motion capture to evaluate movement trajectories during various behaviors under pharmacological treatments in mice. In the current project I aimed to synchronize these highly accurate behavioral recordings with calcium imaging of neuronal activity in cerebellar nuclei. Imaging deep-brain structure activity in a freely behaving mouse remains
an experimental challenge. Here, I aimed to compare the activity of the same neurons across distinct behavioral
tasks to shed light on the neurobiological basis of behavior.
We used adult male C57BL/6 mice in a within-subject, randomized design
(n=6). A high-speed, high-resolution 3D motion capture system (Qualisys, Sweden) was used to track 3D trajectories and velocity of markers during locomotory tasks: novel environment exploration, vertical climbing, 3D exploration, and running on a treadmill increased with each trial up to 40 m/min until failure.  nVoke 2.0 (Inscopix, CA) miniscope was used to image cerebellar interposed nuclei. After correcting the calcium imaging videos
for background and motion noise with Min1pipe, and matching the videos
across recording days (rigid registration on the average image), we
could identify common regions of interest across recordings for a
given mouse (n=20-30, 3 mice).
Preliminary observations indicate that each of the following tasks – exploration of novel 2D and 3D environments, climbing on a mesh wheel, and
treadmill running (20 and 30 m/min) – showed distinct neuronal activity characterizing the
task, which we were able to reproduce in a couple of trials.  While 3D
(mesh wheel and 3D exploration) tasks display varying neural activity
during the recording length, treadmill running displays rather
constant activity.
Further analysis is ongoing to characterize cell-specific activity patterns across tasks and to uncover links between neural
activity and specific behavioral events.

Eiji Uchibe
ATR

Curriculum-Based Transferable Imitation Learning across Robots, Tasks, and Environments

Robots capable of performing a wide range of tasks in diverse environments are essential for addressing labor shortages in areas such as logistics, caregiving, and manufacturing. However, learning general-purpose control policies remains challenging due to differences in robot morphology, environment dynamics, and task objectives. While imitation learning provides a safe, data-driven alternative to reinforcement learning, it often suffers from poor generalization and requires extensive demonstrations tailored to each specific domain. In this work, we propose Universal Entropy-Regularized Imitation Learning (Uni-ERIL), a model-based generative imitation learning framework designed to enable transfer across multiple tasks, environments, and robot embodiments. Uni-ERIL estimates reward functions, policies, and transition dynamics jointly from demonstration data, and uses a curriculum learning approach that incrementally progresses from task-only and environment-only transfer to more complex dual-mode (task and environment) and triple-mode (task, environment, and robot) transfer. Additionally, Uni-ERIL learns an embodiment mapping that aligns representations across robots with different morphologies. We evaluate Uni-ERIL using two dual-arm robots performing seven manipulation tasks across three kitchen environments. Experimental results show that Uni-ERIL consistently outperforms prior methods in both offline learning and online fine-tuning, achieving effective transfer across tasks, environments, and robot platforms. These findings highlight the promise of structured, model-based imitation learning as a scalable approach for acquiring transferable robotic skills in complex real-world settings.

Viktor Zhumatiy

Optimal Stop-Loss Selection via the Kelly Criterion

Determining where to place stop-loss orders remains a long-standing open problem in practical trading. In practice, rules of thumb such as “1% of capital” or “the amount you are willing to lose” are commonly used, even among experienced practitioners.
We propose a formal method for determining optimal stop levels based on a generalized Kelly-criterion framework. The method evaluates a stop-parameterized family of trading policies and selects the policy that maximizes long-run capital growth under a specified level of risk exposure.
The family of policies is constructed by augmenting a trader’s historical empirical trading data with counterfactual outcomes obtained by hypothetically applying stop-loss rules. This produces a dataset of simulated trading trajectories under different stop policies. Applying Kelly-type optimization to these trajectories yields both the optimal capital allocation and a recommended daily stop level.
The framework highlights a perspective that differs from typical reinforcement learning formulations. Many exploration strategies in reinforcement learning operate inside the agent by modifying the policy or action-selection mechanism. In contrast, the Kelly framework evaluates policies at a higher level, treating them as black-box generators of wealth trajectories and deriving risk controls from their observable outcomes. A key feature of this perspective is the explicit treatment of finite resources and survival constraints through capital dynamics. In many biological systems, behavior reflects the interaction of policy adaptation, risk management, and resource allocation processes. Trading environments naturally make these relationships visible, as actions directly affect resource accumulation and survival. Within such a framework, parameters governing exploration and risk-taking arise from capital allocation and survival constraints rather than from heuristic stochastic policy mixing.

Carlos Gutierrez
OIST/Softbank

Sustaining Brain Modeling Platforms in the AI Era: Lessons and Strategies for NeuroWorkflow

We present NeuroWorkflow, a brain modeling platform that integrates AI assistance with human expertise to accelerate model development and workflow construction. Major international initiatives such as EBRAINS, Blue Brain Project, and Allen Brain Institute achieved sustainability through decades of development, combining fundamental research with applied outputs that demonstrate stakeholder value.
Japan's Brain/MINDS 2.0 initiative, launched two years ago, can leverage current AI advances to shorten this traditional development timeline. However, this requires planning for short- to medium-term sustainability from the outset.
This presentation reviews sustainability strategies from established projects, analyzes how AI integration accelerates technical development and user adoption, and discusses approaches for building a digital brain ecosystem in Japan's context.

Jun Igarashi
RIKEN

Connectome-constrained spiking neural network models of the mouse and marmoset cerebral cortex

With the increasing availability of brain data and advances in computational power, data-driven neural network simulations have become increasingly active. However, the integration of diverse datasets, the use of efficient computational methods, and the effective utilization of high-performance computing (HPC) are still limited. To address these challenges, we are promoting the integration of brain data, advanced methods, and HPC for large-scale brain simulations.

In this talk, we will introduce our recent work on data-driven spiking neural network simulations, including oscillatory neural activity in a connectome-constrained spiking neural network model of the mouse and the marmoset cerebral cortex, and the application of data assimilation methods.

Alan Fermin
Hiroshima University

Prospective Role of Heart-to-Brain Dysregulation in the Development of Major Depressive Disorder
 

Ascending visceral and physiological signals inform the brain about bodily homeostasis and survival functions while modulating emotion, cognition, and mental health. In turn, descending brain signals regulate cardiac rhythm, gastrointestinal motility, immune responses, sleep, and sexual arousal. Numerous studies have linked disturbances in these visceral, physiological, and molecular systems to psychiatric disorders, including major depressive disorder (MDD), schizophrenia, Alzheimer's disease, and Parkinson's disease. Nevertheless, the precise mechanisms by which brain-body interactions contribute to the development and persistence of mental disorders, particularly MDD, remain poorly understood.

Here, we present evidence supporting a body-to-brain pathway in the development of MDD. Our findings demonstrate that: (1) structural abnormalities in key brain regions predict elevated depressive symptoms one year later; (2) abnormal insula connectivity linked with reduced body sensibility and depression resilience; (3) reduced heart rate variability (HRV), a marker of autonomic dysregulation, prospectively predicts the transition from euthymia to clinical MDD; (4) ascending heartbeat signals and their ensuing heartbeat-evoked potential (HEP), an electrophysiological index of central cardiac interoception, mediates the relationship between altered insula functional connectivity and concurrent depressive symptoms in healthy adults; and finally, (5) abnormal HRV mediates differences in brain structural integrity between healthy individuals and patients with MDD. These results collectively suggest that visceral signals, particularly those indexed by HRV and HEP, may serve as critical intermediaries in the bidirectional interplay between peripheral physiology and central neural alterations underlying MDD vulnerability and symptomatology.