Research
Research on trustworthy AI, learning from limited data, and uncertainty-aware applications.
I study how AI can learn from limited data and make reliable predictions, especially when a model is uncertain. My work spans evidential and Bayesian deep learning, few-shot learning, and applications in vision, behavioral data, and physical science.
Trustworthy learning
Methods for learning with fewer examples and understanding what a model knows.
Help evidential models keep learning
Generalized the regularization behind our ICML 2023 work to multiple evidential activation functions. The analysis explains when training can stall and how to restore a useful learning signal; experiments span classification, few-shot recognition, and image restoration.
Generalized Regularized Evidential Deep Learning Models ↗More reliable adaptation of vision models
Introduced Bayesian parameter-efficient fine-tuning to address underconfidence when adapting pretrained vision models with few examples. It combines a stronger prior from pretraining with an evidential ensemble for better calibrated uncertainty.
Be Confident in What You Know ↗Learn from every training example
Identified why evidential networks can stop learning from examples that fall in zero-evidence regions. Proposed a regularizer that lets those examples contribute to training, improving performance on challenging datasets.
Learn to Accumulate Evidence from All Training Samples ↗Separate kinds of uncertainty in few-shot prediction
Developed Evidential Conditional Neural Processes to distinguish uncertainty caused by scarce knowledge from uncertainty inherent in the data. The model is designed for few-shot regression and robustness to noisy training tasks.
Evidential Conditional Neural Processes ↗Choose the most useful few-shot tasks
Developed a multidimensional belief measure to estimate uncertainty in meta-learning tasks and guide which tasks to label and train on. A multi-query formulation further reduces labeling and computation in few-shot image classification.
Multidimensional Belief Quantification ↗Research collaborations
Applying uncertainty-aware and physics-informed learning to real-world data.
Learn physical laws from sparse simulations
Developed a physics-informed equation-of-state model that learns an underlying free-energy function from simulated energy and pressure. This construction yields accurate predictions while respecting thermodynamic relationships.
Physics-informed equation-of-state model ↗Predict energy and pressure consistently
Built a deep regression approach for equation-of-state tables with scarce training data, using meta-learning style training, uncertainty regularization, and ensembling. The work also checks thermodynamic consistency of its predictions.
Deep energy-pressure regression ↗Discover patterns in behavior over time
Combined temporal set representations with uncertainty-aware attention to find informative patterns in visual and touch interactions. The study evaluates the approach on child-computer interaction data for autism research.
Deep Temporal Sets with Evidential Reinforced Attentions ↗Find the informative part of a long time series
Developed uncertainty-aware multiple-instance learning to identify the useful segment of a long vessel trajectory, then combine it with satellite imagery when the trajectory prediction is unreliable.
Uncertainty-Aware Multiple Instance Learning ↗Undergraduate research
Redirected Walking in Virtual Reality
Explored how to navigate a larger virtual environment within a smaller physical room. As the main developer, I implemented translation, rotation, and curvature gains in Unity, brought them into an HTC Vive game, and evaluated the experience with participants.
Read the project report ↗ See the project ↗