Adaptive Redox Resistive Memory Programming for Efficient and Robust Class-Incremental Learning.

Resistive memory (RM)-based computing-in-memory (CiM) accelerators provide a promising platform for low-power edge intelligence. However, practical edge AI requires not only efficient inference but also repeated model updates through class-incremental learning (CiL). Implementing CiL on RM substrates exposes a critical material-algorithm mismatch: the write-intensive updates required by CiL are strongly affected by the stochastic programming of filamentary RM devices. Conventional fully programming (FP) mitigates this stochasticity by repeatedly programming and verifying each cell to a precise target conductance, but this exhaustive procedure incurs substantial energy consumption, latency, and device wear across successive CiL stages. Herein, we propose a hardware-aware adaptive programming (AP) strategy that aligns CiL deployment with RM device physics. Microstructural and electrical analyzes reveal that the random spatial distribution of oxygen vacancies gives rise to unavoidable programming variability. Guided by this insight, AP does not attempt to eliminate intrinsic stochasticity through costly compensation. Instead, it updates only the most impactful weights to suppress accuracy loss caused by overall mapping errors, while leaving low-impact weights unchanged. This converts large-scale write-verify operations into targeted updates of a minimal subset of cells, reducing programming overhead without requiring device or material optimization. Validated on a hybrid analog-digital system with a 40 nm, 256 k RM-based CiM core, AP reduces programming energy by 93.0% and programming cycles by more than 90% relative to FP during five-stage CIFAR100 CiL, while achieving a final accuracy of 0.80, close to the 0.81 software baseline. For the more complex ShapeNet 3D point-cloud recognition task across eight stages, AP achieves 92.3% energy savings with only a 0.03 accuracy loss. Moreover, AP improves robustness by reducing programming-error-induced accuracy degradation by 87.7% and 88.4% for the two tasks, respectively. This work bridges algorithmic update requirements and physical programming constraints, enabling robust and energy-efficient lifelong learning on RM-based CiM platforms.