Rapid learning with phase-change memory-based in-memory computing through learning-to-learn.

Area of Science:

• Artificial Intelligence
• Neuromorphic Computing
• Hardware Acceleration

Background:

• Current AI models require extensive resources and data for adaptation, limiting edge applications.
• Human learning demonstrates efficient knowledge transfer and rapid adaptation to new tasks.
• In-memory computing neuromorphic hardware (NMHW) mimics brain principles by co-locating memory and compute.

Purpose of the Study:

• To develop low-power, autonomously learning AI systems capable of rapid adaptation at the edge.
• To integrate learning-to-learn (L2L) principles with in-memory computing neuromorphic hardware (NMHW).
• To demonstrate efficient AI model adaptation using minimal data and computational effort.

Main Methods:

• Paired L2L with NMHW based on phase-change memory devices.
• Implemented AI models on NMHW for real-world task adaptation.
• Utilized meta-training in software for high-precision model preparation.

Main Results:

• Demonstrated AI model versatility in image classification and robotic arm control.
• Achieved rapid learning with few parameter updates on NMHW.
• NMHW-deployed models performed on-par with software equivalents.

Conclusions:

• L2L combined with NMHW enables efficient, rapidly adapting AI for edge applications.
• The proposed approach reduces computational and data requirements for AI model adaptation.
• Software-based meta-training simplifies hardware integration and accuracy concerns.