Linear and Non-linear Dimensionality-Reduction Techniques on Full Hand Kinematics.

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Effective prosthetic hand control requires accurate interpretation of complex hand kinematics.
• Dimensionality reduction techniques are crucial for simplifying high-dimensional kinematic data.
• Comparing linear and non-linear methods is essential for optimizing prosthetic control algorithms.

Purpose of the Study:

• To compare Principal Component Analysis (PCA) and a non-linear Autoencoder Network (nAEN) for representing hand kinematics.
• To determine which method provides a more parsimonious representation for prosthetic hand control.
• To evaluate the effectiveness of nAEN and PCA in capturing essential hand gesture and action characteristics.

Main Methods:

• Hand kinematics data from a wide spectrum of gestures and actions were collected.
• Dimensionality reduction was performed using PCA (linear) and nAEN (non-linear).
• Performance was assessed by data reconstruction accuracy, variance distribution, and movement separability using a linear classifier (SoftMax regression).

Main Results:

• The nAEN demonstrated superior accuracy in reconstructing hand kinematic data from a reduced-dimension latent manifold compared to PCA.
• For two latent dimensions, nAEN captured 94% of data variance, significantly outperforming PCA's 78%.
• The nAEN produced a latent manifold with more separable movements, enabling better distinction between different hand tasks when reconstructed.

Conclusions:

• Non-linear dimensionality reduction using nAEN offers a more effective approach for prosthetic hand control than linear methods like PCA.
• nAEN's ability to capture complex data structures and variance leads to improved representation of hand kinematics.
• This research paves the way for more intuitive and functional prosthetic hand control systems.