DNB: A Joint Learning Framework for Deep Bayesian Nonparametric Clustering.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep neural network-based clustering is vital for image analysis.
• Existing methods often require knowing the number of clusters, which is impractical.
• Unsupervised learning faces challenges with unknown unknowns, including unknown labels and cluster counts.

Purpose of the Study:

• To propose a Bayesian nonparametric framework, Deep Nonparametric Bayes (DNB), for joint image clustering and deep representation learning.
• To address the doubly unsupervised learning problem where both labels and their count are unknown.
• To develop an end-to-end solution for deep clustering that avoids post hoc analysis for cluster number selection.

Main Methods:

• The DNB framework employs an alternating process between forward and backward passes for cluster generation and network learning.
• Utilizes Dirichlet process mixtures to enable partitioning of latent representation space without pre-specifying the number of clusters.
• Offers a principled approach to mitigate the 'trivial solution' problem in deep clustering.

Main Results:

• The proposed DNB method achieves effective clustering performance on benchmark image datasets.
• Demonstrates superior results compared to various other unsupervised image clustering techniques.
• Validates the efficacy of the doubly unsupervised approach in learning both representations and cluster structures.

Conclusions:

• Deep Nonparametric Bayes (DNB) provides a powerful doubly unsupervised solution for image clustering.
• The end-to-end framework effectively learns image clusters and deep representations without prior label information.
• This approach advances unsupervised learning by handling unknown unknowns and addressing the trivial solution problem.