Convolutional Multitimescale Echo State Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Echo State Networks (ESNs) are efficient Recurrent Neural Network (RNN) models for time series.
• Single ESNs struggle to capture multitimescale structures inherent in temporal data.

Purpose of the Study:

• Introduce the Convolutional Multitimescale Echo State Network (ConvMESN).
• Develop a training-efficient model for capturing multitimescale structures and dependencies in temporal data.

Main Methods:

• Construct a multitimescale memory encoder using a multireservoir structure with varying skip lengths.
• Encode time series history into nonlinear multitimescale echo state representations (MESRs).
• Utilize a convolutional layer to learn multiscale temporal dependencies from MESRs.

Main Results:

• MESRs provide demonstrably better discriminative features for time series analysis.
• ConvMESN exhibits efficient memory encoding and strong learning capabilities for complex temporal dependencies.
• Extensive experiments on 18 MTS and 3 action recognition datasets show ConvMESN outperforms existing methods.

Conclusions:

• ConvMESN effectively models temporal data with multitimescale structures.
• The model demonstrates high computational efficiency due to training-free reservoirs and a single convolutional layer.
• ConvMESN advances time series analysis by capturing multitimescale dynamics and outperforming prior approaches.