β-VAE loss
Balances reconstruction quality and latent regularisation through the β parameter.
Core method
MixtureBetaVAE
A mixture-of-encoders β-VAE framework for learning interpretable, traversable latent spaces in heterogeneous scientific data.
Motivation
Scientific data is heterogeneous.
A single encoder can force diverse scientific cases through one representation pathway. This may be limiting when the data contains multiple regimes, such as healthy versus glaucomatous retinal structure, or developing versus non-developing atmospheric systems.
MixtureBetaVAE uses multiple encoders with a shared decoder so that different latent mappings can cooperate while still producing a common reconstruction space. The aim is to support both predictive performance and interpretability.
Architecture
High-level model structure.
Input x
Encoder 1
μ₁, log σ²₁
μ₁, log σ²₁
Encoder 2
μ₂, log σ²₂
μ₂, log σ²₂
Encoder S
μₛ, log σ²ₛ
μₛ, log σ²ₛ
Shared latent space
z
z
Shared decoder
x̂
x̂
Reconstruction
+ optional heads
+ optional heads
Placeholder note:
Later this section can include the exact PyTorch architecture, model diagram, and real model outputs.
Objective
The training objective combines representation, fidelity, and alignment.
Mixture cooperation
Encourages encoder components to cooperate in a shared latent representation space.
Supervised alignment
Optional classification head aligns latent structure with known labels or states.
Adversarial realism
Optional discriminator and feature matching improve local realism in reconstructions.
Domain-aware weighting
Specialised reconstruction weighting can emphasise scientifically important regions.
Centre structure
Optional centre loss can encourage compact class-aware latent regions.
Interactive placeholder
Latent traversal concept.
This demo is currently conceptual. Later it can be connected to actual generated outputs.
Regime A
Transition
Regime B
Suspect-like transition
The latent point is moving through an intermediate region. Later, this can display real RNFL/GCIPL reconstructions or cyclone tensor frames.
Applications
The method is applied across two scientific domains.
Explore how the same representation-learning idea is used for retinal imaging and tropical meteorology.