CV

Doctor of Philosophy

• Teaching assistant of graduate level CS 589 machine learning and CS 651 optimization class
• Research assistant in Information Fusion lab, currently funded by an NIH RO3

Master of Science

Bachelor of Engineering

Audio-Visual Speech Separation via Bottleneck Iterative Network

• Accepted to ICML 2025 Workshop on Machine Learning for Audio
• We work on audio-visual speech separation on noisy audio mixtures of NTCD-TIMIT and LRS3 data set
• We propose a novel multimodal fusion framework on audio and video modality with a bottleneck iterative structure
• Our proposed model outperforms state-of-the-art on SI-SDRi and only requires 50\% of SOTA training time

Longitudinal Multimodal Modeling for Alzheimer’s Early Detection in the Wild

• We introduce information theory-based unsupervised representation learning on brain MRIs as a complement to risk factors, with an estimated mutual information value to evaluate the strength of the dependency to risk factors
• We both train CNN from scratch and finetune foundation models as two representation learning models
• We optimize the procedure of training Alzheimer's forecasting model to include two stages with partial model parameter freezing for stabler validation performance during training process
• We constructed a representative subset from the ADNI dataset aligned with the TADPOLE challenge, consisting of 1,183 patients for training, 169 for validation, and 336 for testing
• We evaluate the forecasting performance using both micro F1-score on CN-MCI-AD forecasting and the precision of capturing MCI to AD transition timing across 100 repeated experiments to ensure statistical significance

Encoding Domain Insights into Multi-modal Fusion: Improved Performance at the Cost of Robustness

• Accepted to ICML 2025 Workshop on Methods and Opportunities at Small Scale
• We work on comparing fusion methods with and without domain knowledge embedded in the model structure
• We conduct experiments on MMSD dataset with both the original task of detecting sarcasm and a synthetic task to control the level of domain knowledge, with additional Gaussian noise added to the inputs to examine robustness
• We find that aligning fusion design with priors boosts clean-data accuracy with limited data but significantly diminishes robustness to noisy inputs

Clustered Vertical Attention for Irregular Time Series Modelling

• Results were submitted to ICML 2019 TimeSeriesWorkshop
• We worked on PhysioNet Challenge 2012 data set
• We improved the prediction accuracy of the in-hospital survival of ICU patients via existing imputation methods
• We ran Minimum Spanning Tree algorithm to determine the most correlated imputed clusters
• We trained separate Attention models for clusters and predicted on a Long Short-term Memory Attention model
• The model got accuracy improvements of 1.5%, 1.3% and 1.8% on 3 different imputation methods