Hyperdimensional computing with holographic and adaptive encoder.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Brain-inspired computing aims to replicate human brain functions in machine learning.
• Hyperdimensional Computing (HDC) uses high-dimensional representations for efficiency and robustness.
• Existing HDC methods use manually selected encoders, limiting task-specific adaptation.

Purpose of the Study:

• To propose FLASH, a novel HDC learning method with an adaptive, learnable encoder.
• To improve overall learning performance and maintain HDC properties.
• To address limitations in current HDC encoder designs.

Main Methods:

• Developed FLASH, a Hyperdimensional Computing (HDC) method.
• Incorporated an adaptive and learnable encoder design.
• Learned the encoder matrix distribution via gradient descent for tailored HDC encoding.

Main Results:

• Tuning the HDC encoder significantly improved accuracy on regression datasets.
• FLASH surpassed current HDC algorithms and RFF-based kernel ridge regression.
• Achieved faster inference speeds compared to baselines.

Conclusions:

• An adaptive encoder is crucial for optimizing Hyperdimensional Computing (HDC).
• Customized high-dimensional representations enhance HDC performance.
• FLASH demonstrates the benefits of learnable encoders in HDC.