Cause of Noise-Free Baseline’s Poor Programming Performance

Determine the exact mechanism responsible for the poor programming performance of the noise-free baseline pulse-time predictor that computes write pulses using the simplified JART valence change mechanism (VCM) memristor model without cycle-to-cycle noise, as compared to the trained neural pulse predictors. Specifically, ascertain why the baseline underperforms despite being derived from a deterministic, noise-free device model, in the context of the switching dynamics observed in the JART VCM model and the evaluation protocol used in this study.

Background

The paper proposes training a neural network to predict voltage pulse durations for programming memristor conductance updates, and compares its performance against a baseline that uses a noise-free device model to compute the pulse time needed to achieve the target conductance change. The baseline is derived by running a deterministic simulation of the simplified JART VCM model (without cycle-to-cycle noise) to obtain pulse times that would produce the same conductance update under idealized conditions.

Empirically, the authors observe that the noise-free baseline exhibits poor performance relative to the trained neural predictor, particularly near conductance boundaries where the JART model shows self-accelerating switching dynamics and S-shaped transfer curves. Although they suggest possible explanations—such as noise assisting switching in ways that bias predictions, or asymmetric noise impacts—the exact cause of the baseline’s poor performance remains unresolved.