Simple Encoder Training for Hyperdimensional Computing

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hyperdimensional computing (HDC) is an emerging, lightweight computing paradigm.
• Data in HDC are represented as high-dimensional vectors and processed using simple algebraic operations.
• Conventional HDC classification trains prototypes while keeping the encoder static.

Purpose of the Study:

• To propose and evaluate a method for training the encoder in a binary/bipolar HDC classification pipeline.
• To leverage native HDC integer and binary operations for encoder training.
• To improve classification accuracy without compromising model size or inference complexity.

Main Methods:

• Developed a novel method for training binary/bipolar HDC encoders.
• Utilized exclusively native HDC integer and binary operations for training.
• Evaluated the trained encoder on standard HDC classification datasets.

Main Results:

• Achieved an average accuracy improvement of 2.13% compared to an untrained encoder.
• Demonstrated performance gains across several common HDC classification datasets.
• Confirmed no increase in model size or inference complexity with the proposed method.

Conclusions:

• Training the HDC encoder using native operations is an effective strategy for enhancing classification accuracy.
• The proposed method offers a practical approach to optimizing HDC classification pipelines.
• This research contributes to the advancement of efficient and accurate hyperdimensional computing models.