Reservoir Computing Approaches for Representation and Classification of Multivariate Time Series

Area of Science:

• Machine Learning
• Time Series Analysis
• Computational Neuroscience

Background:

• Multivariate time series (MTS) classification is crucial across various domains.
• Reservoir computing (RC) offers efficient MTS representation but often lacks accuracy compared to deep learning.
• Existing RC methods struggle to balance speed and classification performance.

Purpose of the Study:

• To introduce an unsupervised approach, the reservoir model space, for learning improved MTS representations using RC.
• To develop a modular RC framework with an open-source Python library for advanced MTS classification architectures.
• To enhance the accuracy and computational efficiency of MTS classifiers based on RC.

Main Methods:

• Developed the reservoir model space, an unsupervised RC method encoding MTS within a linear model predicting reservoir dynamics.
• Implemented an intermediate dimensionality reduction for improved representation quality and computational performance.
• Created a modular RC framework with diverse modules for building and comparing advanced RC architectures for MTS classification.

Main Results:

• The reservoir model space yields superior MTS representations compared to standard RC methods.
• RC classifiers utilizing the proposed representation achieve comparable computational performance with enhanced accuracy.
• Benchmark and real-world MTS datasets demonstrate that the proposed RC classifiers are significantly faster and more accurate than deep learning and time series kernel methods.

Conclusions:

• The reservoir model space and modular RC framework offer a powerful and efficient solution for MTS classification.
• This approach overcomes the traditional accuracy limitations of RC while maintaining its speed advantage.
• The open-source library facilitates the implementation and exploration of advanced RC techniques for MTS analysis.