Back-propagation operation for analog neural network hardware with synapse components having hysteresis...

Area of Science:

• Artificial Intelligence
• Materials Science
• Solid State Physics

Background:

• Analog artificial neural network hardware requires efficient synapse elements.
• Ferroelectric memristors are promising synapse candidates due to their variable resistance.
• Memristor conductance exhibits hysteresis and dispersion, hindering accurate learning algorithms like back-propagation.

Purpose of the Study:

• To propose and simulate a learning operation procedure for analog neural network hardware with ferroelectric memristor synapses.
• To address the challenges of memristor conductance hysteresis and dispersion.
• To enable the application of back-propagation learning in memristor-based neural networks.

Main Methods:

• A weight perturbation technique was used to derive error changes.
• The learning procedure updates pulse voltage based on error reduction, mimicking back-propagation.
• For error increase, pulse voltage is adjusted in the opposite scanning direction to achieve similar conductance, mitigating hysteresis.
• Numerical simulations incorporated conductance dispersion to assess learning probability.

Main Results:

• The proposed procedure successfully eliminated memristor hysteresis, allowing learning operation convergence.
• Simulations incorporating conductance dispersion showed improved learning probability, especially with adequate dispersion magnitude.
• Ferroelectric characteristics, where polarization magnitude is retained with same-polarity voltages, enhanced learning robustness.

Conclusions:

• The developed learning operation procedure is effective for analog neural network hardware utilizing ferroelectric memristor synapses.
• This method provides a viable solution for overcoming inherent memristor non-idealities in AI hardware.
• The findings are significant for the practical implementation of large-scale analog neural networks.