Time series classification with Echo Memory Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Echo State Networks (ESNs) are recurrent neural networks (RNNs) effective for time series prediction.
• Recent applications of ESNs to time series classification (TSC) have limitations, including adapting predictive models or using local time-step classifications, which can reduce accuracy.
• Existing methods fail to effectively select discriminative time series segments, incorporating irrelevant information.

Purpose of the Study:

• To propose a novel end-to-end framework, the Echo Memory Network (EMN), for efficient and accurate time series classification.
• To develop a model that learns both time series dynamics and multi-scale discriminative features.
• To address the limitations of previous ESN-based classifiers by focusing on discriminative patterns.

Main Methods:

• Time series data are projected into a high-dimensional nonlinear space using a reservoir, generating echo states collected in an echo memory matrix.
• A single multi-scale convolutional layer extracts multi-scale features from the echo memory matrix.
• Max-over-time pooling is employed for temporal invariance and selection of key local patterns, followed by a fully-connected layer and softmax for classification.

Main Results:

• The Echo Memory Network (EMN) demonstrates high efficiency with a training-free recurrent layer and a single convolutional layer.
• EMN achieved first-rank performance across 55 time series classification benchmark datasets.
• The model also excelled in four 3D skeleton-based human action recognition tasks, incorporating spatial information fusion strategies.

Conclusions:

• The EMN is a highly efficient and effective end-to-end model for time series classification and human action recognition.
• The proposed architecture successfully learns discriminative features and time series dynamics.
• EMN significantly outperforms previous methods on diverse benchmark datasets.