Sequential Outlier Criterion for Sparsification of Online Adaptive Filtering

Area of Science:

• Machine Learning
• Signal Processing

Background:

• Adaptive filtering methods update knowledge from sequential data over time.
• The learning performance of these systems depends on measuring sample informativeness and subsequent data treatment.

Purpose of the Study:

• To propose a sequential outlier criterion for sparsification in online adaptive filtering.
• To enhance the informativeness measurement of online filtering for more accurate and compact learning networks.

Main Methods:

• A novel method measures sample informativeness based on historical, sequentially adjacent samples.
• Samples are classified as informative, redundant, or abnormal for individual treatment within the learning system.

Main Results:

• The proposed method achieves more effective sample classification.
• Simulations demonstrate more accurate networks in online adaptive filtering tasks.
• Validated on static function estimation, Mackey-Glass, and Lorenz chaotic time series prediction.

Conclusions:

• The sequential outlier criterion enables a more sensible learning process with valid knowledge extraction.
• The method leads to the optimal network configuration in adaptive learning systems.