A Decoder-Free Variational Deep Embedding for Unsupervised Clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Autoencoder (AE)-based deep clustering methods are popular but often retain a redundant decoder post-training.
• The encoder-decoder architecture in deep clustering limits encoder depth and learning capacity.
• Reconstruction error minimization in AEs is theoretically linked to maximizing mutual information (MI).

Purpose of the Study:

• To propose a novel decoder-free variational deep embedding for unsupervised clustering (DFVC).
• To investigate unsupervised representation learning by estimating MI for continuous and categorical representations.
• To develop a more efficient and effective deep clustering framework by removing the decoder.

Main Methods:

• Proposed a decoder-free variational deep embedding for unsupervised clustering (DFVC).
• Utilized mutual information (MI) estimation for continuous and categorical representations, inspired by contrastive self-supervised learning.
• Incorporated data augmentation to enhance discriminative representation learning through MI estimation.
• Constrained the latent space using a Gaussian mixture prior for cluster-friendly embeddings via end-to-end learning.

Main Results:

• The proposed DFVC model effectively learns discriminative representations without a decoder.
• Achieved higher performance compared to various state-of-the-art unsupervised clustering approaches on challenging datasets.
• Demonstrated the theoretical validity of discarding the decoder by focusing on MI maximization.

Conclusions:

• The decoder-free approach significantly improves deep clustering efficiency and effectiveness.
• DFVC offers a promising direction for advanced unsupervised representation learning and clustering.
• The method provides a robust and high-performing alternative to traditional AE-based deep clustering frameworks.