For my PhD project, I set to explore how a nonlinear dimensionality reduction (DR) technique - autoencoder networks (AEs) - can identify low-dimensional latent manifolds of movement data. The thesis focuses on the application of this technique to body machine interfaces (BoMIs). I begin by comparing nonlinear AEs to Principal Component Analysis (PCA), a linear DR method, in capturing essential information of kinematic signals, including hand gestures and object manipulations, as well as electromyographic signals (EMG) obtained from unconstrained movements of shoulders and arms. AEs exhibited higher performance than PCA in the reconstruction of hand kinematic and EMG data from a latent manifold. Therefore, a non-linear DR method has the potential to provide a more effective coding platform for human-machine interfaces (HMIs). I therefore investigated how the choice of hyperparameters (e.g., type of activation function, number of hidden layers, etc.) affected the shape of the latent manifold, particularly its local curvature, and whether these potential effects were correlated with changes in reconstruction performance. To gain more consistent insights on the structure of the nonlinear latent manifold, I developed a visual tool based on classic concepts of cartography. This display offers a direct and intuitive assessment of the AE’s nonlinear transformation. I was able to demonstrate that the cartographic approach makes the visible structure of the latent manifold stable and independent of the AE’s training parameters. As such, the proposed approach is a step toward defining a unique latent manifold. After analyzing the properties of both linear (PCA) and nonlinear (AE) DR techniques, I focused on their applications within the control-scheme of a BoMI. First, I aimed at providing BoMI users with the possibility to switch seamlessly between movement and EMG control. Such approach is essential to utilize the BoMI as a therapeutic tool for promoting recovery of muscle control after neurological injury. In the clinical context it is essential to adapt the operation of the BoMI to the evolving state of its users. My guiding hypothesis for this purpose is that the operation of the interface is facilitated if the BoMI forward map is updated online to match the evolving latent manifold of the user’s motions. Results show that this adaptive approach increased the representational efficiency of the interface and significantly improved users’ task-related performance. As an extension of the application of AE-based BoMI, I developed a non-linear BoMI designed to control an assistive 4D virtual robotic manipulator and tested the interface on a cohort of unimpaired participants, who successfully acquired a high level of robot control. To conclude, I investigated whether AEs can represent and estimate motor learning during the operation of a BoMI. This final study demonstrates that in fact this nonlinear method is effective to accurately track users’ learning process.

Nonlinear dimensionality reduction for human movement analysis with application to body machine interfaces

RIZZOGLIO, FABIO
2021-02-15

Abstract

For my PhD project, I set to explore how a nonlinear dimensionality reduction (DR) technique - autoencoder networks (AEs) - can identify low-dimensional latent manifolds of movement data. The thesis focuses on the application of this technique to body machine interfaces (BoMIs). I begin by comparing nonlinear AEs to Principal Component Analysis (PCA), a linear DR method, in capturing essential information of kinematic signals, including hand gestures and object manipulations, as well as electromyographic signals (EMG) obtained from unconstrained movements of shoulders and arms. AEs exhibited higher performance than PCA in the reconstruction of hand kinematic and EMG data from a latent manifold. Therefore, a non-linear DR method has the potential to provide a more effective coding platform for human-machine interfaces (HMIs). I therefore investigated how the choice of hyperparameters (e.g., type of activation function, number of hidden layers, etc.) affected the shape of the latent manifold, particularly its local curvature, and whether these potential effects were correlated with changes in reconstruction performance. To gain more consistent insights on the structure of the nonlinear latent manifold, I developed a visual tool based on classic concepts of cartography. This display offers a direct and intuitive assessment of the AE’s nonlinear transformation. I was able to demonstrate that the cartographic approach makes the visible structure of the latent manifold stable and independent of the AE’s training parameters. As such, the proposed approach is a step toward defining a unique latent manifold. After analyzing the properties of both linear (PCA) and nonlinear (AE) DR techniques, I focused on their applications within the control-scheme of a BoMI. First, I aimed at providing BoMI users with the possibility to switch seamlessly between movement and EMG control. Such approach is essential to utilize the BoMI as a therapeutic tool for promoting recovery of muscle control after neurological injury. In the clinical context it is essential to adapt the operation of the BoMI to the evolving state of its users. My guiding hypothesis for this purpose is that the operation of the interface is facilitated if the BoMI forward map is updated online to match the evolving latent manifold of the user’s motions. Results show that this adaptive approach increased the representational efficiency of the interface and significantly improved users’ task-related performance. As an extension of the application of AE-based BoMI, I developed a non-linear BoMI designed to control an assistive 4D virtual robotic manipulator and tested the interface on a cohort of unimpaired participants, who successfully acquired a high level of robot control. To conclude, I investigated whether AEs can represent and estimate motor learning during the operation of a BoMI. This final study demonstrates that in fact this nonlinear method is effective to accurately track users’ learning process.
15-feb-2021
File in questo prodotto:
File Dimensione Formato  
phdunige_3835614.pdf

Open Access dal 15/02/2022

Tipologia: Tesi di dottorato
Dimensione 17.18 MB
Formato Adobe PDF
17.18 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11567/1038287
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact