On the Post Hoc Explainability of Optimized Self-Organizing Reservoir Network for Action Recognition

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Human Action Recognition (HAR) is crucial for intelligent systems.
• Echo State Networks (ESNs) offer potential for time-series processing but require careful tuning.
• Unsupervised learning methods are needed for efficient HAR model development.

Purpose of the Study:

• To propose a novel unsupervised network, the Self-Organizing Convolutional Echo State Network (SO-ConvESN), for HAR.
• To leverage Recurrent Plots (RPs) and Recurrence Quantification Analysis (RQA) for reservoir stability and hyperparameter tuning.
• To integrate a Convolutional Neural Network (CNN) for enhanced feature learning and action recognition.

Main Methods:

• Developed the SO-ConvESN architecture for unsupervised learning of ESN parameters.
• Utilized RPs and RQA to analyze and ensure reservoir dynamics stability.
• Cascaded optimized reservoirs with a CNN and employed Hyperparameter Optimization (HPO) for the CNN stage.
• Validated the approach on 3D-skeleton-based action datasets.

Main Results:

• Demonstrated the effectiveness of RPs and RQA in characterizing reservoir dynamics for stable self-organizing reservoirs.
• Showcased the utility of HPO in optimizing the SO-ConvESN for HAR.
• Achieved competitive recognition accuracy on benchmark HAR datasets.
• Validated the SO-ConvESN's ability to learn temporal features and dynamics.

Conclusions:

• The proposed SO-ConvESN provides a stable and effective unsupervised method for HAR.
• RPs, RQA, and HPO are valuable tools for designing and optimizing ESN-based models.
• The SO-ConvESN architecture shows significant promise for advancing human action recognition.