Powering AI at the edge: A robust, memristor-based binarized neural network with near-memory computing and...

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Memristor-based neural networks offer energy-efficient artificial intelligence (AI) and potential for self-powered operation.
• Analog in-memory computing in memristor networks requires stable power, conflicting with unreliable energy harvesters.

Purpose of the Study:

• To develop a robust, energy-efficient memristor-based neural network for AI applications powered by energy harvesters.
• To overcome the limitations of analog computing with unstable power sources by employing a digital near-memory computing approach.

Main Methods:

• Fabrication of a binarized neural network with 32,768 memristors.
• Integration with a miniature wide-bandgap solar cell for power.
• Implementation of a digital near-memory computing architecture with complementary programming and logic-in-sense-amplifier.

Main Results:

• The circuit demonstrated comparable inference performance to a lab bench power supply under high illumination.
• Functional operation was maintained in low illumination, transitioning to approximate computing with slightly reduced accuracy.
• Simulations indicated that misclassifications in low light primarily involved difficult-to-classify images.

Conclusions:

• The developed system lays the groundwork for self-powered AI systems.
• This technology enables the creation of intelligent sensors for health, safety, and environmental monitoring.