Scientific data
OCT-derived retinal maps or atmospheric tensor sequences from meteorological data.
Research
Interpretable representation learning for complex scientific data.
My research develops generative-discriminative models that support prediction, reconstruction, latent traversal, and scientific reasoning.
Research map
One core method, two scientific applications.
Framework
From raw scientific data to interpretable latent spaces.
Representation model
A mixture-of-encoders β-VAE learns latent structure while reconstructing the input.
Prediction
Supervised heads can align latent representations with clinical or physical labels.
Interpretation
Latent traversals and counterfactuals reveal meaningful directions of change.
Evaluation philosophy
Accuracy is necessary, but not sufficient.
Scientific models should be evaluated across multiple axes: prediction, fidelity, interpretability, and stability.
Discrimination
How well does the model separate clinically or physically meaningful states?
Reconstruction
Does the model preserve important scientific structure in generated outputs?
Latent coherence
Do latent traversals correspond to plausible disease or environmental transitions?
Counterfactual stability
Are generated changes stable and meaningful under shifts in domain or cohort?
Research areas
Explore the main research pages.
MixtureBetaVAE
The core method: mixture encoders, shared decoder, latent traversal, and optional supervised alignment.
Open method →Ophthalmic AI
Interpretable representation learning for glaucoma-related retinal structural change.
Open domain →Climate AI
Spatio-temporal tensor learning for tropical cyclone genesis and non-development.
Open domain →Projects
See how the research becomes code.
The projects section includes placeholders for research code, HPC pipelines, OCT preprocessing, and future demos.
View projects