EventHD: Robust and efficient hyperdimensional learning with neuromorphic sensor.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Cognitive Computing

Background:

• Deep learning faces challenges in robustness and energy efficiency.
• Brain-inspired computing, particularly Hyper-Dimensional Computing (HDC), shows promise for enhanced cognitive learning.
• Neuromorphic sensors generate event-based data requiring specialized processing.

Purpose of the Study:

• To present EventHD, an end-to-end learning framework utilizing HDC for efficient and robust learning from neuromorphic sensors.
• To explore HDC as a computational model mimicking brain functionalities for high-efficiency, noise-tolerant computing.
• To enable online learning and cognitive support from raw neuromorphic sensor data.

Main Methods:

• Developed a spatial and temporal encoding scheme to map event-based neuromorphic data into a high-dimensional space.
• Applied HDC mathematics for learning and cognitive tasks, including information association and memorization.
• Integrated a confidence notion for predictions to facilitate self-learning from unlabeled data.

Main Results:

• EventHD operates directly on raw Dynamic Vision Sensor (DVS) data, eliminating costly preprocessing.
• Achieved 14.2x faster processing speeds compared to state-of-the-art learning algorithms.
• Demonstrated 19.8x higher energy efficiency and 5.9x improved computational robustness.

Conclusions:

• EventHD provides an efficient, robust, and noise-tolerant framework for neuromorphic computing using HDC.
• The framework enables online learning and cognitive capabilities directly from event-based sensor data.
• HDC offers a viable alternative to deep learning for specific cognitive tasks, particularly in resource-constrained environments.