Coupled VO₂ Oscillators Circuit as Analog First Layer Filter in Convolutional Neural Networks.

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Traditional computing faces limitations in energy efficiency and speed for AI tasks.
• In-memory computing offers a promising alternative by performing computations where data is stored.
• Phase-change materials like VO2 are being explored for novel computing paradigms.

Purpose of the Study:

• To develop and characterize an in-memory computing platform using coupled VO2 oscillators in a crossbar configuration.
• To demonstrate the neuromorphic computing potential of this platform.
• To apply the platform to a real-world AI task, specifically image recognition.

Main Methods:

• Fabrication of coupled VO2 oscillators in a silicon-based crossbar array.
• Characterization of device variability and reliability.
• Demonstration of neuromorphic computing using oscillator phase relations.
• Integration into a VGG13 convolutional neural network architecture.

Main Results:

• The crossbar configuration significantly improves area density and oscillation frequency.
• The devices exhibit low variability and high reliability, enabling experiments with 4-coupled oscillators.
• Successful demonstration of neuromorphic computing capabilities.
• Achieved 95% accuracy on the MNIST dataset when replacing digital filtering with oscillating circuits.

Conclusions:

• The proposed in-memory computing platform based on coupled VO2 oscillators offers a promising approach for efficient AI hardware.
• The crossbar architecture provides advantages in scalability and performance.
• This work validates the potential of using oscillatory dynamics for computational tasks in neural networks.