Electrode area precision and odd-state quantization failure in neuromorphic memristive synapses
Applied Physics Letters · 16.11.2026 · 129 (20)
DOI: 10.1063/5.0348077 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Physics Letters, online erschienen am 19.11.2026. Verfasst von Kameyab Raza Abidi und 1 weiteren Autor:innen. Die Arbeit nennt 35 Literaturangaben.
Abstract (Originalsprache)
In interface-type memristive synapses, both on-state and off-state conductances scale linearly with electrode area, while cycle-to-cycle variability scales as the square root of area, so the number of distinguishable states grows as the square root of electrode area, making area a direct precision parameter. We derive a closed-form precision-area law, calibrate it to thin-film manganite (GCMO and PCMO) device data, and apply it to neural-network classification accuracy. Unexpectedly, post-mapping trained weights onto three conductance levels yields only 20.5% accuracy, near random for ten classes, whereas two levels yield 63.8%. The collapse traces to a dead-weight mechanism, where a quantizer with an odd number of levels places one level at zero, and for networks with large weight outlier ratios, the fraction of silenced weights approaches unity, a failure invisible to the signal-to-quantization-noise ratio, which incorrectly ranks three conductance levels above two. Hardware-aware training with a straight-through estimator eliminates the collapse, recovering 86.1% on average at three conductance levels, reaching within 0.3% points of the ideal network accuracy at six levels, and requiring 4.8-fold smaller electrode area than max-range post-mapping and 3.2 times less than post-mapping with the matched quantization range. Uncompensated device-to-device variability imposes a practical resolvability ceiling on the number of usable states, and halving the device-to-device variability coefficient from 0.10 to 0.05 gains more accuracy at eight conductance levels than doubling the electrode area. These results provide a design framework that translates electrode area and device variability into achievable classification accuracy.
Zitieren
APA
Abidi, K. R., & Das, G. M. (2026). Electrode area precision and odd-state quantization failure in neuromorphic memristive synapses. Applied Physics Letters, 129(20). https://doi.org/10.1063/5.0348077
Vancouver
Abidi KR, Das GM. Electrode area precision and odd-state quantization failure in neuromorphic memristive synapses. Applied Physics Letters. 2026;129(20). doi:10.1063/5.0348077
BibTeX
@article{Abidi2026,
author = {Abidi, Kameyab Raza and Das, Gour Mohan},
title = {Electrode area precision and odd-state quantization failure in neuromorphic memristive synapses},
journal = {Applied Physics Letters},
year = {2026},
volume = {129},
number = {20},
doi = {10.1063/5.0348077}
}
EUVIMED