Dimensionality Reduction for Prosthetic Hand Control

Portnova-Fahreeva, Alexandra

doi:https://doi.org/10.21985/n2-ap90-9q91

Work

Dimensionality Reduction for Prosthetic Hand Control

Public

Download PDF

Download All Files (.zip)

In this document, I demonstrate that: 1) Linear basis functions cannot outperform nonlinear ones to represent hand kinematics 2) Nonlinear autoencoders outperform PCA on the dimensionality reduction of hand kinematics, 3) Nonlinear autoencoders outperform PCA in human gait representation and recurrent nonlinear autoencoders can seamlessly express the temporal dynamics, 4) Factors that aid and inhibit one’s learning to operate low-dimensional controllers of high-dimensional hand systems, 5) Factors that are important for myoelectric latent representations for low-dimensional control.Due to the nonlinear nature of hand kinematics during object grasping and gesturing, linear methods, such as Principal Component Analysis (PCA), cannot outperform their nonlinear counterparts as claimed in Yan et al. (Yan et al., 2020). Here I present a demonstration of this by applying a simple three-layer nonlinear AE network to Yan’s dataset and highlight the superiority of the network over PCA. I present an analysis of the nonlinear AE network in reducing the dimension of complex hand kinematics and human gait, confirming the superiority of the nonlinear AE structure in its ability to efficiently compress biological data and maintain an equal spread of variance across latent dimensions. I also show that an AE network with a temporal component can perform a more accurate movement classification and individual identification. Next, I present a low-dimensional myoelectric controller, in which a high-dimensional virtual hand with 17 DOFs is controlled via a 2D space with muscle signals from the wrist. I conduct three studies to understand which factors affect how the user learns to operate such a controller. In particular, short exploration times are insufficient to facilitate learning. Inhibition on the learning process also happens due to the difficulty of operating a myoelectric interface. Lastly, it is important to provide the users with a clear link between the underlying low-dimensional (2D) controller and the high-dimensional (17D) task of controlling a virtual hand to accelerate learning and achieve the most optimal performances at the end of the training. After identifying the challenges of learning to operate a low-dimensional controller, I explore what makes a latent space of myoelectric signals useful in the context of low-dimensional controllers. The final study demonstrates that the latent space structure greatly dependents on both the input data parameters and the dimensionality-reduction method type.

Creator