An encoding framework for binarized images using hyperdimensional computing.

Area of Science:

• Machine Learning
• Computer Science
• Artificial Intelligence

Background:

• Hyperdimensional Computing (HDC) is a brain-inspired, lightweight machine learning technique.
• HDC is suitable for edge computing, wearable devices, and on-device AI due to its low computational complexity.
• Effective data encoding into the hyperdimensional space is crucial for HDC performance.

Purpose of the Study:

• To propose a novel, lightweight encoding approach for binarized images using native HD arithmetic.
• To preserve spatial similarity in binarized images through point of interest selection and local linear mapping.
• To enhance the performance and robustness of Hyperdimensional Computing for image classification.

Main Methods:

• The proposed method utilizes native HD arithmetic vector operations for encoding.
• Point of interest selection and local linear mapping are employed to preserve spatial relationships.
• The approach focuses on encoding binarized images.

Main Results:

• Achieved 97.92% accuracy on the MNIST dataset and 84.62% on Fashion-MNIST.
• Outperformed existing native HDC encoding methods and matched hybrid HDC and binarized neural networks.
• Demonstrated superior robustness to noise and blur compared to baseline encoding methods.

Conclusions:

• The novel encoding approach offers competitive performance for binarized image classification using HDC.
• This method provides a lightweight and robust alternative for edge AI applications.
• The technique advances the application of HDC in resource-constrained environments.