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

Area of Science:

• Materials Science and Engineering
• Computer Engineering
• Artificial Intelligence

Background:

• Resistive memory (RM)-based computing-in-memory (CiM) offers low-power edge intelligence.
• Class-incremental learning (CiL) is crucial for practical edge AI but faces challenges with RM device stochasticity and write-intensive updates.
• Conventional fully programming (FP) mitigates stochasticity but leads to high energy consumption, latency, and device wear.

Purpose of the Study:

• To develop a hardware-aware adaptive programming (AP) strategy for CiL on RM substrates.
• To align CiL implementation with the physical constraints of RM devices, specifically programming variability.
• To reduce the overhead associated with repeated model updates in edge AI systems.

Main Methods:

• Proposed a hardware-aware adaptive programming (AP) strategy that targets only the most impactful weights for updates.
• Leveraged microstructural and electrical analyses to understand the impact of oxygen vacancy distribution on programming variability.
• Validated the AP strategy on a hybrid analog-digital system with a 40 nm, 256 k RM-based CiM core using CIFAR100 and ShapeNet datasets.

Main Results:

• AP reduced programming energy by 93.0% and programming cycles by over 90% compared to FP during CIFAR100 CiL, achieving 0.80 accuracy.
• For ShapeNet, AP achieved 92.3% energy savings with only a 0.03 accuracy loss across eight learning stages.
• AP improved robustness, reducing programming-error-induced accuracy degradation by up to 88.4%.

Conclusions:

• The adaptive programming strategy effectively bridges algorithmic update requirements with physical programming constraints in RM devices.
• AP enables robust and energy-efficient lifelong learning on RM-based CiM platforms without requiring material or device modifications.
• This approach significantly enhances the practicality of edge AI by optimizing model updates for resource-constrained environments.