Filament-Free Bulk RRAM with High Endurance and Long Retention for Neuromorphic Few-Shot Learning On-Chip

Bulk switching RRAM devices have emerged to address nonidealities of filamentary RRAM for AI at the edge. However, long retention and high endurance needed for continual on-chip learning have yet to be demonstrated. Here, we report a filament-free and multilevel bulk RRAM (b-RRAM) technology with long retention and high endurance. The oxygen vacancy gradient across the $\text{TiO}_{\mathrm{x}}$ switching layer and $\text{Al}_{2}\mathrm{O}_{3}$ tunnel barrier are designed to suppress filament formation and achieve multilevel switching through modulation of oxygen vacancy distribution. AI-Ti-O oxygen barrier enables reliable switching with high uniformity, retention and endurance. We develop a compact model that captures both DC and pulse switching. $\mathrm{M}\Omega$-level resistance and switching current nonlinearity allow highly accurate read/write operations and matrix vector multiplications (MVM) in selectorless b-RRAM crossbars. A neuro-inspired few-shot learning (FSL) algorithm based on dendritic computation and behavioral time plasticity (BTSP) is mapped onto b-RRAM crossbar arrays. Weights are continuously updated on-chip to learn shortest navigation to reward within a few trials in a 2D maze. We also present FSL of large-size images with hyperdimensional (HD) computing using our b-RRAM model. System-level simulations with 64-level b-RRAM arrays show an increase in accuracy by 5.9 % and 5.4 % for a 5-way 10-shot learning of CIFAR-100 and CUB-200 datasets as compared to the state-of-the-art FSL-HD.