Multi-Energy Quasi-Symplectic Langevin Inference for Latent Disentangled Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Variational autoencoders (VAEs) are common for large datasets.
• 3D image VAEs struggle with disentangled representations, low-variance Evidence Lower Bounds (ELBO), and model size.
• Existing methods face computational bottlenecks in Langevin-based flow inference.

Purpose of the Study:

• To develop an efficient inference framework for 3D image modeling.
• To achieve unsupervised disentanglement of appearance and morphology features.
• To create a lightweight VAE with low-variance ELBO.

Main Methods:

• Proposed a Langevin dynamics-based inference framework integrating target data information.
• Employed multi-scale energy-level encoding for unsupervised disentanglement of appearance and morphology.
• Utilized a quasi-symplectic integrator to mitigate Hessian-related computational bottlenecks.

Main Results:

• Demonstrated theoretical and empirical effectiveness compared to existing methods.
• Achieved high-quality generation on public benchmarks and clinical 3D imaging datasets.
• Learned disentangled shape and appearance representations with a 1.7M parameter model.

Conclusions:

• Langevin-VAE offers an efficient and effective solution for 3D image modeling.
• The framework successfully disentangles features and maintains a lightweight model.
• This approach advances VAE capabilities for complex 3D data analysis.