Collective dynamics and long-range order in thermal neuristor networks.

Area of Science:

• Neuromorphic Engineering
• Condensed Matter Physics
• Computational Neuroscience

Background:

• Scalable and energy-efficient neuromorphic devices are crucial for advanced AI.
• Spiking oscillators that mimic biological neurons are key components.
• Vanadium dioxide (VO2) resistive memories offer unique thermal properties.

Purpose of the Study:

• To introduce and characterize a new class of spiking oscillators: thermal neuristors.
• To investigate the collective dynamics and phase structure of thermal neuristor networks.
• To evaluate the computational capabilities of thermal neuristor arrays in tasks like image recognition and time series prediction.

Main Methods:

• Fabrication and characterization of thermal neuristor devices based on VO2.
• Experimental and theoretical analysis of network dynamics under varying thermal coupling and input voltage.
• Implementation of reservoir computing using thermal neuristor arrays for benchmark tasks.

Main Results:

• Demonstrated a rich phase structure in thermal neuristor networks, tunable by thermal coupling and input voltage.
• Identified novel phases with long-range order not arising from criticality, but from time non-local responses.
• Achieved high accuracy in image recognition and time series prediction using thermal neuristor arrays via reservoir computing.

Conclusions:

• Thermal neuristors represent a promising platform for scalable and energy-efficient neuromorphic computing.
• The findings challenge the necessity of criticality for efficient neuromorphic computation, highlighting the role of non-local dynamics.
• This research offers insights into brain function and the design principles of artificial neural systems.