Metaplasticity-Enabled Graphene Quantum Dot Devices for Mitigating Catastrophic Forgetting in Artificial Neural...

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Artificial Intelligence

Background:

• Deep neural networks struggle with continuous learning due to simplified models of biological neural circuits.
• Catastrophic forgetting limits the ability of artificial intelligence to retain new information while learning.

Purpose of the Study:

• To introduce artificial synaptic devices with graphene quantum dots (GQDs) that exhibit metaplasticity.
• To demonstrate a hardware solution for overcoming catastrophic forgetting in deep neural networks.

Main Methods:

• Developed artificial synaptic devices using graphene quantum dots (GQDs) to achieve metaplasticity.
• Utilized interface-mediated modifications in asymmetric conductive pathways for synaptic plasticity.
• Implemented metaplasticity principles in deep neural networks for enhanced generalization.

Main Results:

• The GQD-enhanced devices demonstrated metaplasticity, regulating memory stability and learning plasticity.
• Achieved 97% accuracy on a sequential task (fourth MNIST dataset) while maintaining over 94% on previous tasks.
• Successfully addressed catastrophic forgetting in deep neural networks.

Conclusions:

• Metaplasticity is crucial for generalization in deep neural networks, enabling fluid adaptation to new data.
• GQD-based artificial synaptic devices offer a viable hardware approach for neuromorphic systems.
• This research bridges the gap between artificial and biological neural network learning capabilities.