DGVAC: Disentangled Gaussian-vMF alignment for deep clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Unsupervised deep clustering faces challenges in learning discriminative and robust representations.
• Variational autoencoder (VAE)-based models suffer from representation entanglement, hindering clustering performance.
• Disentangling discrete class identity from continuous variations is crucial for effective clustering.

Purpose of the Study:

• To propose a disentangled variational framework for improved unsupervised deep clustering.
• To address the problem of representation entanglement in VAE-based clustering models.
• To enhance clustering alignment and robustness to intra-class variations.

Main Methods:

• Decomposing the latent space into Euclidean and hyperspherical subspaces with Gaussian and von Mises-Fisher priors, respectively.
• Employing a probabilistic alignment strategy integrating Product-of-Experts (PoE) with self-supervised learning for robust subspace learning.
• Developing a staged Evidence Lower Bound (ELBO) optimization scheme with four synergistic components: reconstruction, clustering, categorical regularization, and Gaussian regularization.

Main Results:

• The proposed method consistently improves clustering accuracy across benchmark datasets.
• Enhanced generative quality is achieved through disentangled representation learning.
• The framework demonstrates robustness to perturbations, enabling stable latent representations for clustering.

Conclusions:

• The disentangled variational framework effectively addresses representation entanglement in deep clustering.
• The integration of heterogeneous latent subspaces and probabilistic alignment significantly boosts clustering performance.
• The method offers a promising direction for advancing unsupervised deep clustering through disentangled representation learning.