PDGMM-VAE: A Variational Autoencoder with Adaptive Per-Dimension Gaussian Mixture Model Priors for Nonlinear ICA

Researchers Yuan-Hao Wei and Yan-Jie Sun have introduced PDGMM-VAE, a novel variational autoencoder architecture designed to solve the challenging problem of nonlinear Independent Component Analysis (ICA). The core innovation lies in its source-oriented design: each latent dimension, explicitly representing an individual source signal, is equipped with its own Gaussian Mixture Model (GMM) prior. This is a significant departure from conventional VAEs, which use a single, simple prior (like a standard Gaussian) shared across all latent dimensions. By imposing these per-dimension, heterogeneous priors, the model can capture the diverse and often non-Gaussian statistical properties of real-world source signals, which is crucial for effective separation.

Critically, the parameters of these individual GMM priors are not fixed or pre-defined. Instead, they are adaptively learned and refined end-to-end alongside the encoder and decoder parameters during training. In this framework, the encoder learns to act as a demixing function, mapping observed mixtures back to the estimated independent sources, while the decoder learns to reconstruct the original observation from these separated components. This creates a fully probabilistic, trainable system for blind source separation. The authors' experiments demonstrate that PDGMM-VAE successfully recovers latent sources and achieves strong separation performance on both linear and, more importantly, nonlinear mixing problems, validating the power of its adaptive, source-specific prior modeling.