Deep clustering with a Dynamic Autoencoder: From reconstruction towards centroids construction.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Unsupervised learning lacks direct cost functions for variations and similarities.
• Static objective functions in unsupervised learning miss opportunities presented by smooth system dynamics.
• Integrating smooth dynamics into training can leverage gradual, uncertain knowledge from pseudo-supervision.

Purpose of the Study:

• To propose a novel deep clustering model that addresses the reconstruction-clustering trade-off.
• To introduce a dynamic objective function for unsupervised learning.
• To enhance the utilization of pseudo-supervision in clustering tasks.

Main Methods:

• Developed Dynamic Autoencoder (DynAE), a novel deep clustering model.
• Implemented a dynamic objective function that gradually shifts from reconstruction to construction.
• Evaluated the model on benchmark datasets using unsupervised learning principles.

Main Results:

• DynAE achieves state-of-the-art results in deep clustering.
• The dynamic objective function effectively balances reconstruction and clustering objectives.
• The model demonstrates superior performance compared to existing deep clustering methods.

Conclusions:

• Dynamic Autoencoder (DynAE) offers a significant advancement in deep clustering.
• Dynamically adjusting objective functions is a promising direction for unsupervised learning.
• The proposed method effectively handles the trade-off between reconstruction and clustering.