Lipschitz-based robustness estimation for hyperdimensional learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Machine learning models require robustness evaluation for practical applications.
• Hyperdimensional computing (HDC) offers a neurosymbolic approach but lacks robustness analysis.
• Input perturbations pose a significant challenge to HDC model reliability.

Purpose of the Study:

• To develop a theoretical framework for evaluating HDC classifier robustness against input perturbations.
• To propose a method for enhancing HDC model robustness based on the developed framework.
• To quantify the impact of noise on HDC model predictions.

Main Methods:

• Proposed a novel theoretical framework to assess hyperdimensional classifier robustness.
• Developed a robustness measure providing an upper bound for tolerable noise magnitude.
• Introduced methods to calculate robustness based on dataset and hyperdimensional encoding.
• Implemented an optimization scheme varying Gaussian distribution variance for hypervector encoding.

Main Results:

• The proposed measure provides a theoretical upper bound on noise tolerance for HDC models.
• The optimization scheme successfully increased the average robustness of HDC models.
• Model accuracy was maintained while enhancing robustness.
• The effectiveness of the robustness measure and enhancement method was practically demonstrated.

Conclusions:

• The study presents a significant advancement in understanding and improving HDC model robustness.
• The developed framework and methods offer practical tools for building more reliable HDC systems.
• Future work can explore diverse datasets and encoding strategies to further validate the approach.