Self-organizing kernel-based convolutional echo state network for human action recognition / Lee Gin Chong
The research works on three-dimensional (3D)-skeleton-joints-based human action recognition (HAR) remains demanding. Among the approaches, Echo State Networks (ESNs) are a reservoir computing method that considers skeleton joints human actions as multivariate time series and attempts to identify and...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2022
|
| Subjects: | |
| Online Access: | http://studentsrepo.um.edu.my/14464/1/Lee_Gin_Chong.pdf http://studentsrepo.um.edu.my/14464/2/Lee_Gin_Chong.pdf http://studentsrepo.um.edu.my/14464/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The research works on three-dimensional (3D)-skeleton-joints-based human action recognition (HAR) remains demanding. Among the approaches, Echo State Networks (ESNs) are a reservoir computing method that considers skeleton joints human actions as multivariate time series and attempts to identify and model the dynamical temporal features in 3D space. Despite the random initialization of the ESN's input and reservoir weights may reduce the computational cost, this may raise instability and variance in generalization and hence diminish reproducibility. Moreover, ESN remains a black-box algorithm. Notably, it lacks explainability consideration to understand the input-dependent reservoir dynamics, specifically while configuring the optimal hyperparameters for HAR. Besides, following the body of Convolutional Echo State Network (ConvESN) work to incorporate modeling dynamics and multiscale temporal feature in handling HAR problem, the model may be very sensitive to the selection of hyperparameters in Convolutional Neural Network (CNN). This work addresses these issues by proposing a novel approach known as Self-Organizing Convolutional Echo State Network (SO-ConvESN) for HAR. Specifically, this work proposes an unsupervised self-organizing network for learning node centroids and interconnectivity maps compatible with the deterministic initialization of ESN reservoir weights. To ensure stability and echo state property (ESP) in the self-organizing reservoir, this work further exploits the Recurrent Plots (RPs) and Recurrence Quantification Analysis (RQA) techniques for explainability and characterization of the dynamics of the self-organizing reservoir and hence tuning two critical ESN hyperparameters: input scaling ( |
|---|