Answers are grounded in this paper and cite their sources.
We demonstrate a simple 1R self‐rectifying halide perovskite memristor with dynamically tunable thresholds for selector‐free crossbar arrays and programmable T‐ReLU mapping. The memristors achieve an ON/OFF ratio > 10 5 and rectification ratio > 10 4 . The switching mechanism involves two synergistic processes: bulk ion migration within CsCuCl 3 provides and sustains the low‐resistance state essential for a high ON/OFF ratio, while the Ag‐Cl‐containing oxidized interfacial layer introduces a Schottky barrier and trap filling/release mechanism that governs the HRS to LRS switching. Systematic simulations using CNN architectures reveal a dataset‐dependent threshold sensitivity: MNIST accuracy remains robust (> 98%) regardless of threshold variations, whereas CIFAR‐10 accuracy drops sharply as positive thresholds increase. Based on this complexity‐dependent trend, we exploited the bi‐directional tunability for dual‐mode applications: (i) Negative threshold operation in CNNs maintains 91.0% accuracy for low‐light image recognition (comparable to the ReLU baseline); (ii) Positive threshold modulation in a customized SparseResNet‐18 architecture is explored for sparsity‐aware edge computing, where thresholds τ ∈ [0, 0.3] are adjusted according to simulated battery levels to control the sparsity‐accuracy trade‐off. Under low‐battery conditions (τ = 0.3), the simulated network achieves 76.3% activation sparsity while maintaining 90.1% accuracy without retraining. This work highlights the potential of tunable activation thresholds for reconfigurable neuromorphic systems.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: