PhD Researcher · Computational Neuroscience · KU Leuven
Liuyin Yang
I build brain foundation models spanning ECoG, EEG, sEEG and MEG for brain–computer interfaces — teaching large models pretrained on scalp EEG to transfer into intracranial recordings, and turning neural signals into real-time control, auditory, and motor decoding.
Background
Curriculum Vitae
Education
-
2021 — nowPhD, Biomedical Sciences (Medicine)KU Leuven · Leuven, BelgiumTopic: EEG-based brain–computer interfaces. Advisor: Prof. Marc M. Van Hulle.
-
2020 — 21Advanced Master, Artificial IntelligenceKU Leuven · Big Data AnalyticsCum laude
-
2019 — 20MSc, Electronics & ICT EngineeringKU Leuven · Engineering TechnologyCum laude
-
2017 — 19BSc, Electronics EngineeringKU LeuvenCum laude
-
2015 — 17BEng, Electrical Engineering & AutomationChongqing University · China
Research Experience
-
2021 — nowPhD ResearcherComputational Neuroscience Group, KU LeuvenEEG foundation models for BCI; cross-modality scalp-to-intracranial transfer (CORTEG, NeurIPS 2026); auditory decoding (ICASSP 2023 — 2nd place); winner of the EEG Foundation Challenge at NeurIPS 2025; online motor-imagery BCI systems.
-
2019 — 20Master's Thesis ResearcherKU Leuven · Engineering TechnologyReal-time adaptive-filter speech enhancement on a DSP board, deployed in an IoT communication system.
Skills
15 papers · 138 citations
Publications
As of November 2026 · Source: Google Scholar
-
2026
NeurIPS 2026 · Poster
CORTEG: Foundation Models Enable Cross-Modality Representation Transfer from Scalp to Intracranial Brain Recordings
Can a large model pretrained on scalp EEG be adapted to intracranial (ECoG) recordings? CORTEG bridges the two with an electrode-aware spatial adapter and a dual-stream tokenizer for low-frequency and high-gamma activity, then calibrates to a held-out patient in minutes — matching or beating task-specific baselines on finger-trajectory and audio-envelope regression.
Conference on Neural Information Processing Systems (NeurIPS) · Poster · arXiv:2605.10337
-
2026
ICLR 2026 · Poster
Are EEG Foundation Models Worth It? Comparative Evaluation with Traditional Decoders in Diverse BCI Tasks
Do EEG foundation models actually beat traditional decoders? A statistically rigorous benchmark of state-of-the-art EEG foundation models across diverse datasets, decoding tasks and six evaluation protocols, alongside ST-EEGFormer, a simple ViT baseline pretrained with masked autoencoding on over 8M EEG segments. Fine-tuned foundation models do well in data-rich, population-level settings, but often fail to significantly outperform compact networks or even classical non-neural decoders when data are scarce.
International Conference on Learning Representations (ICLR) · Poster
33 citations Scholar -
2026
arXiv 2026 · Preprint
Characterization of Speech Imagery in Scalp EEG and Comparison with Motor Imagery
Does imagined speech look like motor imagery in scalp EEG? In 34 participants, finger motor imagery showed the expected contralateral mu/alpha and beta desynchronization, whereas speech imagery showed a weaker, more distributed increase in alpha power relative to no-task trials, and was harder to decode (balanced accuracy 0.56 vs. 0.72 for motor imagery).
arXiv preprint · arXiv:2606.12223
-
2026
ICLR 2026 · TSALM Workshop · Poster
Large EEG Foundation Model Learns Informative Low-Frequency Representations from Intracranial Brain Signals
Can a scalp-EEG foundation model transfer to invasive recordings? An ECoG-to-EEG channel-projection module and a lightweight adaptation strategy let the pretrained backbone decode ECoG, and the adapted model outperforms conventional ECoG decoders at extracting finger-movement information from low-frequency (128 Hz) signals.
1st ICLR Workshop on Time Series in the Age of Large Models (TSALM) · Poster
1 citation Scholar -
2026
ICLR 2026 · TSALM Workshop · Poster
Do We Need Domain-Specific Time-Series Models? Insights from EEG Classification Benchmarks
Does the broad pre-training of generic time-series foundation models pay off in a specialized domain? Benchmarking Mantis (~8M parameters) and MOMENT (~300M) on four public EEG datasets against EEG foundation models and classic baselines: fine-tuning often matches EEG-specific models while linear probing transfers poorly, and a randomly initialized Mantis performs comparably to the pretrained one.
1st ICLR Workshop on Time Series in the Age of Large Models (TSALM) · Poster
-
2026
ICLR 2026 · TSALM Workshop
Learning Robust EEG Representations with a Large Spatiotemporal Transformer as a Foundation Model
Introduces ST-EEGFormer, a simple spatiotemporal EEG transformer that projects patches of raw EEG into an embedding space enriched with spatial and temporal embeddings, so that one model can handle recordings with different channel setups, and evaluates how well it adapts across diverse BCI control paradigms.
1st ICLR Workshop on Time Series in the Age of Large Models (TSALM)
1 citation OpenReview -
2026
KU Leuven RDR 2026 · Open dataset
A Multi-Paradigm EEG Dataset for Multi-Task BCI Applications
EEG from 34 healthy participants, each recorded in a single session covering SSVEP, motor imagery and imagined speech, plus structured artifact recordings (eye blinks, jaw clenching, breath holding, hand clenching), resting state and questionnaires. Randomly interleaved test blocks enable realistic asynchronous decoding, and standardized preprocessing pipelines support benchmarking and cross-paradigm transfer learning.
KU Leuven Research Data Repository · doi:10.48804/3EUSTS
-
2025
arXiv 2025 · Preprint
CoMET: A Contrastive-Masked Brain Foundation Model for Universal EEG Representation
Masked reconstruction alone over-focuses on local similarity between highly correlated neighboring channels. CoMET pairs an EEG masked autoencoder with a contrastive framework using mirror-scale augmentation to learn global discriminative patterns; pretrained on over one million samples from more than 3,000 subjects, it reaches state-of-the-art results on ten downstream datasets.
arXiv preprint · arXiv:2509.00314
11 citations arXiv -
2025
ICORR 2025
On the Impact of Proprioception in EEG Representations and Decoding during Human-Hand Exoskeleton Interaction
How does the proprioceptive feedback of a hand exoskeleton shape EEG? In 25 participants performing imagined, passive and combined imagined-plus-passive finger flexion with a cable-driven exoskeleton, exoskeleton-driven movement evoked significantly stronger sensorimotor alpha/beta desynchronization than imagery alone and was readily decoded against rest, but separating motor intention from the exoskeleton’s own feedback remained hard.
International Conference on Rehabilitation Robotics (ICORR) · pp. 186–192
1 citation DOI -
2024
JNER 2024
Unraveling EEG Correlates of Unimanual Finger Movements: Insights from Non-Repetitive Flexion and Extension Tasks
Do individual finger movements have distinct, decodable EEG correlates? Sixteen participants performed five individual and four coordinated finger flexions and extensions during high-density EEG and finger tracking. Movement versus rest was detected with over 80% accuracy, but telling fingers apart was hard because their neural correlates overlap, with the thumb standing out as the most distinctive.
Journal of NeuroEngineering and Rehabilitation · vol. 21, art. 228
23 citations DOI -
2024
J. Neural Eng. 2024
Binocularly Incongruent, Multifrequency-Coded SSVEP in VR: Feasibility and Characteristics
A systematic study of binocularly incongruent SSVEP in virtual reality, where each eye sees a stimulus flickering at a different frequency, covering dual-, triple- and quadruple-frequency codes (several of them new) and comparing them with congruent stimulation. The results show binocular summation and suppression effects, and underline how paradigm choice and decoder design shape performance and user comfort.
Journal of Neural Engineering · vol. 21(5), 056013
9 citations DOI -
2024
IEEE OJSP 2024
Sea-Wave: Speech Envelope Reconstruction from Auditory EEG with an Adapted WaveNet
Extends our ICASSP 2023 challenge entry into Sea-Wave, a WaveNet-based model that reconstructs the speech envelope from EEG and outperforms the previous state of the art. Ablations and a hyperparameter search give a subject-independent Pearson correlation of 0.226 on seen and 0.116 on unseen subjects, and subject-specific fine-tuning adds a further 30% average relative gain on seen subjects.
IEEE Open Journal of Signal Processing · vol. 5, pp. 686–699
6 citations DOI -
2023
ICASSP 2023 · 2nd place
Decoding Auditory EEG Responses Using an Adapted WaveNet
A WaveNet-based model that reconstructs the speech envelope from auditory EEG. It placed second in the regression subtask of the ICASSP 2023 Auditory EEG Challenge and achieved the highest score on held-out subjects, with only 341k trainable parameters.
IEEE ICASSP 2023 — 2nd place, Auditory EEG Grand Challenge · pp. 1–2
19 citations DOI -
2023
Sensors 2023
Real-Time Navigation in Google Street View Using a Motor Imagery-Based BCI
A motor-imagery BCI for navigating Google Street View with only eight dry EEG electrodes. A middle-level control scheme keeps the number of commands small and eye blinks act as an extra control signal to avoid decoder mistakes; offline and online experiments with 20 participants showed acceptable performance despite the minimal setup.
Sensors · vol. 23(3), 1704
18 citations DOI -
2021
Biosensors 2021
Chronic Study on Brainwave Authentication in a Real-Life Setting: An LSTM-Based Bagging Approach
Can brainwaves authenticate users outside the lab? Using a commercial dry-electrode headset and chronic recordings from 15 people performing and imagining motor tasks, an LSTM with bootstrap aggregating (bagging) reached 93.0% authentication accuracy with a 1.9% false-acceptance rate on the combined tasks.
Biosensors · vol. 11(10), 404
15 citations DOI
No publications of this type yet.
Invited · teaching · service
Talks, Teaching & Service
Invited & Selected Talks
-
Dec 2025Spotlight Talk — EEG Foundation Challenge 2025NeurIPS 2025 Competition Track · San Diego
-
2025EEG Foundation ModelsIEEE NER 2025 Workshop
-
2025Scaling EEG Foundation Model PretrainingKU Leuven / UHasselt Tier-2 HPC User Day
Teaching & Supervision
-
recurringTeaching AssistantBrain-Computer Interfaces (MSc course) · KU Leuven
-
2024 — 25Master's Thesis Daily SupervisorKU LeuvenVR-based imagined-speech speller
Brain connectivity with a brain foundation model -
2023 — 24Master's Thesis Daily SupervisorKU LeuvenVR-based binocular SSVEP system
Peer Review
-
journalsJournal ReviewerJournal of Neural Engineering (JNE) · IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)
-
2026Conference ReviewerICLR 2027 · ICML 2026 Workshop FMSD
Funding · 72,119 GPU hours on VSC
Grants & Fellowships
-
2026VSC Compute Project 2026_064Flemish Supercomputer Center (VSC) · HPC GPU allocation for intracranial (ECoG) foundation-model training39,961 GPU hours
-
2025VSC Compute Project 2025_500Flemish Supercomputer Center (VSC) · Special competition grant for the EEG Foundation Challenge (NeurIPS 2025)7,497 GPU hours
-
2025VSC Compute Project 2025_038Flemish Supercomputer Center (VSC) · HPC GPU allocation · vscentrum.be24,661 GPU hours
-
2021 — 25FWO PhD Fellowship Strategic Basic Research (SB)Research Foundation – Flanders (FWO) · fwo.be
Recognition
Awards & Honors
-
20251st Place — EEG Foundation Challenge 2025From Cross-Task to Cross-Subject EEG Decoding · NeurIPS 2025 Competition Track · eeg2025.github.io
-
20232nd Place — ICASSP 2023 Grand ChallengeAuditory EEG Decoding Challenge · WaveNet regression model
-
2019 — 21Cum laude — three consecutive KU Leuven degreesAdvanced Master AI (2021) · MSc Electronics & ICT (2020) · BSc Electronics (2019)
Let's talk
Contact
Elsewhere