Research staff · Suzhou, China
I work at the intersection of biomedical signal processing and machine learning, with a focus on EEG decoding and brain–computer interfaces. I am interested in learning reliable representations from noisy, high-dimensional signals when data are limited and vary across people and recording conditions.
I am preparing to apply for PhD programmes for Fall 2027.
My work with multichannel EEG has led me to a broader question: how can we learn reliable structure from high-dimensional time series with limited observations and substantial variability? I want to explore principled signal-processing and machine-learning methods that improve the robustness and generalization of neural and physiological signal models.
Research Staff · July 2026–present
BCI & Embodied Intelligence Division, East China Institute of Optoelectronic Integrated Devices
I contribute to EEG acquisition, data standardization, quality control, and model-development workflows. My work includes implementing and evaluating data-processing and training pipelines, investigating cross-task adaptation, and studying robustness under heterogeneous data conditions.
I design and implement EEG experimental paradigms for a hospital-partnered clinical study, including Go/No-Go tasks, stimulus presentation, trial structure, and behavioral measures. I also support ethics review, participant procedures, EEG acquisition, and preliminary analysis of responses to neurostimulation.
I develop Windows host software for an EEG acquisition device, covering device communication, data reception and parsing, real-time waveform visualization, channel selection, and device-status controls.
Research Assistant · Supervised by Prof. Dejan Markovic (UCLA) · Wuhan, China
I developed a leakage-controlled EEG decoding pipeline for BCI Competition IV Dataset 2a, using 22-channel recordings from nine participants. Controlled EEGNet ablations compared temporal averaging with high-resolution and position-aware residual attention, showing that most within-subject gains arose from temporal aggregation rather than learned attention alone.
I extended the study to cross-subject generalization with nested leave-one-subject-out evaluation. Transductive unsupervised Euclidean Alignment, using only unlabeled EEG from the held-out target participant, improved mean accuracy from 41.63% to 50.88% and Macro-F1 from 33.89% to 48.55%.
Preprocessing included artifact-trial rejection, 4–40 Hz band-pass filtering, temporal resampling, and fold-specific normalization estimated exclusively from training data. Subject-specific stratified four-fold evaluation used nested validation and early stopping.
Temporal averaging improved mean within-subject accuracy/Macro-F1 from 54.51%/52.33% to 60.14%/58.06%; the best attention variant reached 60.91%/59.03%.
For cross-subject evaluation, inner GroupKFold model selection was restricted to source participants. The source-only baseline achieved 41.63% accuracy and 33.89% Macro-F1. Alignment used unlabeled target EEG without target labels and improved results for all nine participants; paired subject-level testing was significant (two-sided Wilcoxon, p = 0.0039).
Wuhan University · Advisor: Prof. Yong Liu
I developed a device-aware MNIST classification framework to study how weight discretization and device fluctuations affect neural-network performance. I modeled discrete conductance states and approximately 10% resistance fluctuations informed by the group’s NbOx measurements.
Comparisons of ideal, binary, unipolar, and bipolar models showed how device non-idealities reduce recognition accuracy and stability. I controlled hidden-layer size and training iterations and evaluated accuracy curves, residuals, and confusion matrices on 10,000 test samples.
Kong, L. et al.
2023 International Conference on Medical Imaging, Sanitation, and Biological Pharmacy
M.S. in Electrical and Computer Engineering · Boston, MA
Selected coursework: Deep Learning, Introduction to Learning from Data, Computational Optical Imaging, and Cloud Computing.
B.S. in Physics · Wuhan, China
Training in mathematical physics, probability and statistics, computational physics, and programming.
I built a Webots simulation of a wheelchair-mounted Neuronics IPR arm, implementing inverse-kinematics-based reaching, grasping, and object transfer. A modular interface combined keyboard commands, gaze-based input, and preprogrammed motions.
I developed a workflow translating cloud security requirements and more than 100 AWS rules into executable OPA/Rego policies. The system used structured LLM extraction, Python, AWS Lambda, Step Functions, S3, and GitHub Actions for reproducible testing.
I welcome conversations about PhD opportunities in neural signal processing, brain–computer interfaces, and machine learning.
Suzhou, China · Curriculum vitae (PDF)