How do neural connections become a mind?

Overview

Unraveling the mysteries of the human brain requires a bold leap beyond traditional neuroscience. Inspired by the transformative success of Large Language Models (LLMs), the Neuro-X project proposes the development of a Large Brain Model (LBM) — a unified AI system pre-trained on massive, multi-modal datasets of brain structure, activity, and behavior.

Neuro-X, echoing ambitious scientific endeavors like Project Apollo, seeks to decode the neural syntax of the mind, leveraging the brain’s own self-organized dynamic principles to predict and interpret mental processes.

Inside-Out Framework & Temporal Syntax

Drawing upon György Buzsáki’s “inside-out” neuroscientific framework, Neuro-X embraces the brain’s action-driven nature: viewing it as a prediction machine that continually generates internal action plans and learns from external consequences.

LBMs leverage Spatiotemporal Brain Transformers and State-Space Models (e.g. Mamba) to capture neural rhythms and self-organized trajectories across hierarchical spatial regions, providing a temporal syntax for inter-regional communication.

Key Frontiers

1. Dynamic Functional Connectomics: Moving beyond static correlation matrices by learning time-varying attention patterns that capture fleeting cognitive states, emotional transitions, and subtle neuropathological shifts.
2. Multi-Modal Brain Transformers: Seamlessly fusing functional MRI (high spatial resolution), EEG (high temporal resolution), diffusion MRI (tractography structural scaffolds), and polygenic scores into a joint latent space.
3. Decodable Representations: Zero-shot and few-shot task adaptation for cognitive state decoding, psychiatric risk stratification, and generative neural simulation.

Groundbreaking Architectures for 4D Brain Signals

Biological neural signals possess unique spatial symmetries, continuous temporal flows, and non-Euclidean geometries that challenge conventional computer vision and NLP architectures. Connectome Lab pioneers customized deep learning paradigms designed from first neuroscientific principles:

1. NeuroMamba: State-Space Foundation Model for fMRI

• Authors: Jubin Choi et al. (NeurIPS 2025 Foundation Models for Brain and Body - Spotlight)
• Core Innovation: Overcomes the quadratic computational complexity of Transformers in handling long continuous 4D fMRI scans by integrating selective State-Space Models (Mamba). Enables efficient linear-time context modeling across tens of thousands of volumetric brain voxels over thousands of time frames.

2. DIVER-0: Fully Channel-Equivariant EEG Foundation Model

• Authors: Dongyeop Han, Ahhyun Lee, Taeyang Lee, Sebin Lee et al. (ICML 2025 GenBio - Spotlight)
• Core Innovation: Solves the notorious electrode montage mismatch problem across clinical and research EEG datasets by ensuring strict spatial equivariance over arbitrary channel layouts.

3. Spatiotemporal Multi-Band Dynamics

• Authors: Sangyoon Bae et al. (arXiv:2503.23394)
• Core Innovation: Novel frequency-specific multi-band attention mechanisms that disentangle canonical neural oscillation bands within low-frequency hemodynamic fluctuations in fMRI.

Unraveling Complex Gene-Brain-Behavior Dynamics

Psychiatric conditions such as major depressive disorder, anxiety, and neurodevelopmental conditions arise from intricate interplays between polygenic risk architectures and environmental stressors across developmental timepoints.

High-Impact Discoveries

1. Polygenic Architecture in Children (Nature Communications, 2025):
   • Eunji Lee, Bogyeom Kim, Jiook Cha et al. demonstrated how distributed polygenic risk scores manifest in macro-scale cortical morphology, functional connectome reconfigurations, and behavioral phenotypes across thousands of children in the ABCD Study.
2. Multigenerational Polygenic Mediation (Molecular Psychiatry, 2025):
   • Investigated how parental and ancestral psychiatric histories transmit neurobiological vulnerability to offspring through specific polygenic risk profiles and altered white matter microstructure.
3. White-Matter Informed Deep Learning for Youth Depression (Communications Medicine, 2026):
   • Jungyoun Janice Min, Yoonjung Yoonie Joo, Jiook Cha et al. built diffusion MRI-informed deep learning that integrates microstructural white matter measures with genome-wide PRS, improving prediction of depression in adolescents.

Why quantum, and what actually blocks it

Quantum machine learning is usually sold on asymptotics. The wall we actually hit is more prosaic: variational circuits on current hardware are noisy, small, and hard to train. Barren plateaus flatten gradients as circuits widen, device noise biases every expectation value, and the parameter-shift rule makes on-hardware training cost climb with every parameter added.

Our work, much of it with Brookhaven National Laboratory, runs at that wall along three threads: make circuits trainable at scale, make them fit real brain signals, and get them onto real devices.

1. Trainable at scale — multi-chip ensembles

• Authors: Junghoon Justin Park, Jiook Cha, Samuel Yen-Chi Chen, Huan-Hsin Tseng, Shinjae Yoo — arXiv:2505.08782
• Idea: partition a high-dimensional computation across an ensemble of smaller, independently operating quantum chips, entangling only at controlled inter-chip boundaries.
• Result: mitigates barren plateaus, improves generalization, and — unusually — reduces quantum error bias and variance simultaneously, with no separate error-mitigation pass. Validated on MNIST, FashionMNIST, CIFAR-10, and a real-world PhysioNet EEG dataset.
• The same ensemble idea carries into reinforcement learning in It’s-A-Me, Quantum Mario (IEEE QAI 2025), which follows Over the Quantum Rainbow (IEEE QCE 2024), where variational circuits were paired with a Rainbow DQN agent and then explained rather than left as a black box.

2. Fitting real brain signals

• Quantum Time-series Transformer — Junghoon Justin Park, Jungwoo Seo, Sangyoon Bae, Samuel Yen-Chi Chen, Huan-Hsin Tseng, Jiook Cha, Shinjae Yoo (IEEE QCE 2025). Classical self-attention costs quadratic time and a parameter budget that neuroimaging cohorts cannot always feed. Building attention out of a Linear Combination of Unitaries and Quantum Singular Value Transformation brings that to polylogarithmic complexity, and holds up with fewer parameters and smaller samples. Evaluated on resting-state fMRI from ABCD and the UK Biobank.
• Q-DIVER — Junghoon Justin Park, YeongHyeon Park, Jiook Cha (IEEE QCNC 2026). A differentiable quantum classifier sits on top of the lab’s pretrained DIVER-1 EEG encoder, and Differentiable Quantum Architecture Search discovers the circuit topology during end-to-end fine-tuning rather than fixing an ansatz by hand. On PhysioNet Motor Imagery it matches a classical MLP head (test F1 63.49%) using roughly 50× fewer task-specific parameters (2.10M vs 105.02M) — a budget a portable BCI could actually carry.
• HQTCN — Junghoon Justin Park, Maria Pak, Sebin Lee, Samuel Yen-Chi Chen, Shinjae Yoo, Huan-Hsin Tseng, Jiook Cha (IEEE QCNC 2026). A hybrid quantum temporal convolutional network: dilated temporal windows are sampled into shared quantum circuits, so one circuit reads several time scales without a parameter count that grows with the sequence.

3. Running on real devices

• Image Classification on IBM Quantum Computers — Junghoon Justin Park, Jiook Cha, Jun-gyeong Park, Hwidong Yoo, Kwangmin Yu — arXiv:2607.17705. Ten-class MNIST, trained and served end-to-end on a 127-qubit IBM Eagle processor. A two-phase protocol separates gradient-based classical optimization of the encoder and readout from gradient-free optimization of the quantum parameters, removing the parameter-shift cost that makes on-hardware training impractical. It is also the first use of Quantum Multi-Programming on a trained classifier: several copies of the circuit are packed onto one device for parallel inference at no cost in mean accuracy.
• QPATE (ICASSP 2024) — with William H. Watkins, Heehwan Wang, and Sangyoon Bae — carries differential-privacy guarantees into quantum classifiers through private aggregation of teacher ensembles.

Where this is going

Brain data is where the scaling pressure is real. A 4D fMRI scan or a long EEG montage is exactly the high-dimensional, long-sequence regime in which classical attention gets expensive and quantum encodings start to look interesting.

None of the results above claims a quantum advantage on brain data today. The claim is narrower and more useful: circuits that train, that fit the signal, and that run.

We recruit undergraduate research interns on this line year-round. Prior knowledge of quantum mechanics is not required.

The hard end of affect

Most affective neuroscience runs on stimuli chosen for experimental control: static images, isolated faces, single adjectives on a valence scale. The feelings people actually care about do not arrive that way. Awe, being moved by a piece of music, the particular colour a memory takes on when you revisit it — these are mixed, temporally extended, and bound to context.

This axis studies affect at that harder end, with three moves: measure the feelings that resist a single scale, model the media that evoke them, and sometimes put the whole apparatus in a room with an audience.

1. Awe is ambivalent

• Authors: Jinwoo Yi, Dong Yeop Han, Seung-Yeop Oh, Jiook Cha — Communications Psychology (2025)
• Awe is usually filed under positive emotion. Across behavior and cortical responses, this work finds it is better described as ambivalent — carrying positive and negative affect at once, rather than sitting at one end of a valence axis.
• That matters beyond awe: it is a concrete case where the standard one-dimensional valence model loses information that the brain evidently keeps.

2. Reconstructing a remembered feeling

• Revisiting Your Memory: Reconstruction of Affect-Contextualized Memory via EEG-guided Audiovisual Generation — Joonwoo Kwon, Heehwan Wang, Jinwoo Yi, Sooyoung Kim, Shinjae Yoo, Yuewei Lin, Jiook Cha (ACM Multimedia workshop, 2025).
• Recorded EEG steers a generative audiovisual system, so a recalled episode is re-rendered with the affective colouring the recall actually carried — not the content of the memory, but its feeling-tone.

3. Modeling aesthetic style

• AesFA: An Aesthetic Feature-Aware Arbitrary Neural Style Transfer — Joonwoo Kwon, Soo Young Kim, Yuewei Lin, Shinjae Yoo, Jiook Cha (AAAI 2024). Style transfer that separates aesthetic features by frequency rather than leaning on a heavy pretrained encoder.
• Repurposing Image Diffusion Models for Training-Free Music Style Transfer on Mel-Spectrograms — Heehwan Wang, Joonwoo Kwon, Sooyoung Kim, Jungwoo Seo, Shinjae Yoo, Yuewei Lin, Jiook Cha (IEEE ICIP 2026). An image diffusion model, pointed at a mel-spectrogram, transfers musical style with no training at all.
• These are not side projects. A generative model of style is a working hypothesis about what a stimulus does to a listener or a viewer — which is exactly what an affect experiment needs to manipulate.

4. Naturalistic settings

Film, music, and virtual reality are how this axis gets ecological validity. The lab’s earlier work on anticipation of high-arousal film clips (Social Cognitive and Affective Neuroscience, 2014) established the approach, and it now runs through VR protocols for anxiety and panic disorder shared with the computational-psychiatry axis.

And then, sometimes, a gallery

“Study the science of art. Study the art of science. Develop your senses — especially learn how to see. Realize that everything connects to everything else.” — Leonardo da Vinci

In late 2025, Seokjin Moon, Heehwan Wang, and Kyungjin Oh took this apparatus out of the lab and into OB/Scene Focus (옵/신 포커스) as an interactive exhibition, 「Connectome: Reconstruction of Memory」. Wireless high-density EEG (Enobio) and real-time decoding turned visitors’ affective states into generative soundscapes and projections as they stood there.

The exhibition is an output of this research, not its definition — but it is a good test of it. A model of feeling that cannot survive contact with an audience is probably not modeling much.