Training a Hyperdimensional Computing Classifier Using a Threshold on Its Confidence

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Hyperdimensional computing (HDC) offers efficient machine learning for resource-constrained devices like wearables.
• Current HDC training primarily focuses on misclassified samples, potentially missing opportunities for model refinement.
• HDC's computational efficiency makes it suitable for edge computing and Internet-of-Things (IoT) applications.

Purpose of the Study:

• To improve Hyperdimensional Computing (HDC) classification accuracy and confidence.
• To introduce an extended training procedure that incorporates low-confidence correctly classified samples.
• To evaluate the effectiveness of the proposed training method across diverse datasets.

Main Methods:

• An extended HDC training procedure was developed, incorporating samples classified with low confidence.
• A tunable confidence threshold was introduced to optimize classification accuracy for different datasets.
• The proposed method was evaluated on UCIHAR, CTG, ISOLET, and HAND datasets.

Main Results:

• The extended training procedure consistently improved classification performance compared to baseline HDC methods.
• Performance gains were observed across various confidence threshold values on tested datasets.
• The model demonstrated increased confidence in correctly classifying samples after the enhanced training.

Conclusions:

• The proposed training extension enhances HDC model accuracy and prediction confidence.
• Incorporating low-confidence correct predictions is a viable strategy for improving HDC performance.
• This approach offers a more robust and reliable HDC classifier for edge AI applications.