Bayesian neural network with unified entropy source and synapse weights using 3D 16-layer Fe-diode array

Area of Science:

• Materials Science
• Artificial Intelligence
• Device Physics

Background:

• Edge AI systems demand high-frequency operation, challenging traditional entropy sources that degrade with frequency.
• Existing entropy sources struggle with the stability required for edge AI under varying temperatures and frequencies.

Purpose of the Study:

• To investigate the suitability of Fe-diode devices as a stable, high-frequency entropy source for edge AI.
• To experimentally implement a Bayesian neural network using Fe-diode devices on silicon-based chips.

Main Methods:

• Characterization of Fe-diode device noise density across varying read voltages, frequencies, and temperatures.
• Experimental implementation of a hierarchical Bayesian neural network utilizing a 3D 16-layer Fe-diode array.
• Demonstration of a unified entropy source and 4-state synapse architecture.

Main Results:

• Fe-diode devices exhibit stable noise density, modifiable by read voltage, across high frequencies and temperature fluctuations.
• The implemented Bayesian neural network achieved high recognition accuracy.
• The system demonstrated high area efficiency and a wide working temperature range.

Conclusions:

• Fe-diode devices are physically suitable for edge AI applications requiring high frequencies and environmental stability.
• The developed Fe-diode-based Bayesian neural network offers low-energy in-situ training and high performance for edge AI.