Unsupervised Continual Learning in Streaming Environments

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Deep clustering networks (DCNs) excel at feature extraction but require manual engineering.
• Automatic DCN construction in streaming data is challenging due to high labeling costs.
• Unsupervised approaches are increasingly demanded for data stream analysis.

Purpose of the Study:

• To present an unsupervised method for constructing deep clustering networks (DCNs) on the fly for data streams.
• To introduce the autonomous deep clustering network (ADCN) that combines deep learning and clustering.
• To address the need for automated network construction and feature extraction in unsupervised streaming environments.

Main Methods:

• Developed an autonomous deep clustering network (ADCN) integrating feature extraction and autonomous fully connected layers.
• Implemented self-evolving network depth and width based on bias-variance decomposition of reconstruction loss.
• Incorporated self-clustering in deep embedding spaces and latent-based regularization to prevent catastrophic forgetting.

Main Results:

• ADCN demonstrates superior performance compared to existing methods in rigorous numerical studies.
• The network autonomously constructs its structure within streaming environments.
• ADCN operates effectively without requiring labeled samples for model updates.

Conclusions:

• ADCN offers a fully autonomous solution for deep clustering network construction in streaming data.
• The unsupervised approach bypasses the need for feature engineering and expensive data labeling.
• The proposed method advances the field of online machine learning and deep clustering.