Nonlinear memristor model with exact solution allows for ex situ reservoir computing training and in situ inference

Area of Science:

• Neuromorphic engineering
• Physical reservoir computing
• Materials science

Background:

• Memristive devices offer low-power, high-dimensional mapping for complex computations.
• Ion-channel memristors exhibit rich dynamics suitable for biological edge applications.
• Accurate memristor modeling is crucial for optimizing reservoir computing hyperparameters ex situ.

Purpose of the Study:

• To develop a generalized sigmoidal growth model for ion-channel memristor conductance.
• To evaluate the model's accuracy against linear and logistic models.
• To apply the model for ex situ optimization and in situ testing of a handwritten digit classification task.

Main Methods:

• A generalized sigmoidal growth model for ion-channel memristor conductance was derived and solved analytically.
• The model's predictive accuracy was compared with linear and logistic models.
• The sigmoidal model was used to optimize and train a physical reservoir computing network ex situ for MNIST classification.

Main Results:

• The sigmoidal model demonstrated superior accuracy in predicting memristor dynamics compared to linear and logistic models.
• The ex situ optimized and in situ tested network achieved 90.6% accuracy on the MNIST handwritten digit classification task.
• The proposed modeling approach enables efficient ex situ hyperparameter tuning, saving time and energy.

Conclusions:

• The generalized sigmoidal model provides a more accurate representation of ion-channel memristor dynamics.
• This accurate modeling facilitates effective ex situ training of memristive physical reservoir computing systems.
• The approach shows significant potential for efficient and accurate edge computing applications, particularly in biological settings.