Cause of Poor Performance of Noise-Free Baseline

Determine the exact cause of the poor performance of the noise-free baseline pulse predictor that computes pulse time using a noise-free memristor model (unaffected by cycle-to-cycle noise) to achieve a desired conductance change ΔG, when applied to conductance programming with the JART valence change mechanism (VCM) memristor model. Ascertain whether the self-accelerating switching dynamics of SET/RESET and asymmetric noise effects are responsible for the observed degradation.

Background

The paper evaluates neural networks that predict programming pulse times for memristors modeled using the JART VCM model, and compares them against a baseline that uses a noise-free model to compute pulse duration for a desired conductance change. In simulation, the trained neural predictors outperform the noise-free baseline across much of the operating range.

In the Discussion section, the authors explicitly state that the noise-free baseline performs poorly and that the exact reason is unclear. They hypothesize that self-accelerating switching dynamics and asymmetric noise impacts may introduce biases that degrade the baseline’s effectiveness, but the definitive cause remains unresolved.