Huiguang He

Research Overview

Huiguang He focuses on neural computation, brain-computer interfaces, and multimodal neural signal decoding. His trajectory spans from early medical image analysis and structural brain network modeling to advanced generative models for reconstructing sensory perception from brain activity. Currently, his research emphasizes integrating large language models and multimodal fusion techniques to decode complex visual, auditory, and emotional states.

Research Trajectory

His early research concentrated on medical image processing and the structural analysis of brain networks. He contributed to foundational imaging techniques, such as those detailed in A New Fast Surface Generation Scheme and Its Application in Medical Image. Building on this structural focus, he participated in developing kernel-based methods for clinical diagnosis in Accurate prediction of AD patients using cortical thickness networks. This line of inquiry naturally extended into the research direction of Multimodal Fusion of Structural and Functional Brain Networks, as his group increasingly integrated diverse data types for clinical applications, demonstrated by the development of Multi-modal multiple kernel learning for accurate identification of Tourette syndrome children.

By the late 2010s, his focus expanded significantly into neural decoding and brain-computer interfaces (BCIs). A major theme became the reconstruction of visual stimuli from functional magnetic resonance imaging (fMRI). In Reconstructing Perceived Images From Human Brain Activities With Bayesian Deep Multiview Learning, he introduced a deep generative multiview model to predict external visual information from evoked brain activities, contributing to the broader area of Shared Latent Spaces for Multimodal Data Integration. Simultaneously, he addressed the high individual variability in electroencephalogram (EEG) signals. To reduce the need for extensive personal calibration data, he proposed methods in Multisource Transfer Learning for Cross-Subject EEG Emotion Recognition, advancing the research direction of Cross-Subject Generalization for Neural Decoding. His work in BCI paradigms also included foundational data contributions, such as releasing an open-access dataset in Multi-channel EEG recording during motor imagery of different joints from the same limb.

As neural decoding tasks grew more complex, he explored sophisticated graph-theoretic and multimodal alignment techniques. He addressed fundamental limitations in message-passing paradigms, such as information bottlenecks, in Growing Like a Tree: Finding Trunks From Graph Skeleton Trees, which conceptually intersects with structural methodologies in the area of Skeletonized Alignment for Robust White Matter Microstructure Comparison. Returning to visual decoding, he tackled the challenge of generalizing to novel visual categories without paired neural training data. In Decoding Visual Neural Representations by Multimodal Learning of Brain-Visual-Linguistic Features, he leveraged semantic knowledge to bridge this gap, a key advancement in the research direction of Multimodal Alignment for Neural Signal Decoding. He further refined these representations in Interpretable Visual Neural Decoding with Unsupervised Semantic Disentanglement, projecting neural activities onto feature spaces where dimensions correspond to distinct visual attributes.

In recent years, his group has heavily integrated large language models (LLMs) and diffusion models into neural decoding pipelines. For auditory processing, he developed BP-GPT: Auditory Neural Decoding Using fMRI-prompted LLM, utilizing brain representations as prompts for text generation, and Reverse the auditory processing pathway: Coarse-to-fine audio reconstruction from human brain activity, which employs latent diffusion models to reconstruct high-fidelity audio. His research into affective computing has also deepened. He mapped the brain's emotional architecture in Identifying the hierarchical emotional areas in the human brain through information fusion and developed incremental learning frameworks for complex emotional states in EmoGrowth: Incremental Multi-label Emotion Decoding with Augmented Emotional Relation Graph. To better capture the high-dimensional geometry of emotion, he utilized AI models as cognitive agents in Bridging the behavior-neural gap: A multimodal AI reveals the brain's geometry of emotion more accurately than human self-reports.

Parallel to fMRI decoding, he continues to optimize real-time, interactive EEG systems, contributing to the area of Low Latency Neural Decoding for Interactive Brain Interfaces. To enable rapid deployment of Rapid Serial Visual Presentation (RSVP) BCIs, he integrated language-image priors to eliminate the need for target-specific calibration in Integrating Language-Image Prior into EEG Decoding for Cross-Task Zero-Calibration RSVP-BCI, and combined EEG with eye-tracking data in Exploring EEG and eye movement fusion for multi-class target RSVP-BCI. He has also explored hardware interventions, demonstrating that targeted repetitive transcranial magnetic stimulation (rTMS) can boost SSVEP BCI performance in Enhancing visual brain-computer interface through V1-targeted RTMS by modulating visual attention. Culminating this trajectory of multimodal integration, he introduced Mind-Omni: A Unified Multi-Task Framework for Brain-Vision-Language Modeling via Discrete Diffusion, a comprehensive model that converts continuous brain signals into discrete tokens to jointly process fMRI, vision, and language across diverse encoding and decoding tasks.

Current Directions

The most recent papers in the evidence indicate a strong focus on unified, multimodal architectures for neural decoding. Recent work investigates the use of discrete diffusion and large language models to create generalist frameworks capable of handling simultaneous brain, vision, and language tasks. Additionally, ongoing research addresses the practical deployment of interactive brain-computer interfaces, specifically developing zero-calibration methods for RSVP-BCIs by fusing EEG with eye-tracking and language-image priors. Another active area involves mapping the high-dimensional geometry of human emotion, utilizing AI models as cognitive agents to bridge the gap between neural activity and behavioral self-reports.