In-memory factorization of holographic perceptual representations

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Materials Science

Background:

• Disentangling sensory signal attributes is crucial for perception, cognition, and AI development.
• Current methods face challenges with high-dimensional data and computational complexity.

Purpose of the Study:

• To present a compute engine for efficient factorization of high-dimensional holographic representations.
• To leverage hyperdimensional computing and analogue in-memory computing for enhanced factorization capabilities.

Main Methods:

• Utilized computation-in-superposition from brain-inspired hyperdimensional computing.
• Employed analogue in-memory computing with nanoscale memristive devices.
• Developed an iterative in-memory factorizer architecture.

Main Results:

• Demonstrated factorization of problems five orders of magnitude larger than previously possible.
• Significantly reduced computational time and space complexity.
• Achieved constant time for matrix-vector multiplication operations, independent of matrix size.
• Successfully factorized visual perceptual representations experimentally.

Conclusions:

• The developed compute engine offers a powerful solution for complex sensory data factorization.
• This approach drastically improves efficiency in terms of scale, time, and space complexity.
• The technology holds promise for advancing artificial intelligence and cognitive systems.