Mixture encoders
Separate encoder components produce latent statistics for different representation pathways.
Project · Research model
MixtureBetaVAE
A mixture-of-encoders β-VAE framework for interpretable representation learning across heterogeneous scientific domains.
Summary
Why this project exists.
MixtureBetaVAE is designed for scientific settings where the goal is not merely to classify examples, but to learn latent representations that can be inspected, traversed, and interpreted.
The model uses multiple encoders and a shared decoder to support heterogeneous data regimes. This is useful when different examples may follow different structural pathways but still need to live in a common latent and generative space.
The current website version uses placeholder visuals and descriptions. Later, this page can include training curves, architecture diagrams, ablation summaries, generated outputs, and links to code or papers.
Technical structure
Core building blocks.
Shared decoder
A common decoder encourages a unified generative space across encoder components.
β-VAE objective
Balances latent regularisation with reconstruction fidelity through rate-distortion control.
Optional supervised head
Aligns latent representations with class labels while preserving generative structure.
Adversarial realism
Optional discriminator and feature matching can improve local quality in generated outputs.
Latent traversal
Supports counterfactual exploration across clinically or physically meaningful transitions.
Placeholder outputs
Ready for real model results.
Latent space plot
Placeholder for UMAP, PCA, t-SNE, or learned latent trajectory visualisation.
OCT reconstruction
Placeholder for RNFL/GCIPL input, reconstruction, and traversal comparison.
Cyclone tensor traversal
Placeholder for meteorological variable frames or animated tensor outputs.
What I learned
Project reflection placeholders.
Scientific ML needs more than high AUC.
The practical value of the model depends on whether the learned representation supports meaningful scientific questions.
Generative fidelity and discrimination can compete.
Loss design should be treated carefully because improving one objective can weaken another.
Interpretability must be validated, not assumed.
Latent traversals only matter if their changes are clinically or physically plausible.