Stochastic yet Precise: Multi-Level RRAM Crossbar Arrays for In-Memory Learning, Security, and Generative AI

Emerging resistive memory technologies provide a unique opportunity to unify computation, storage, security, and data generation within a single hardware substrate. This talk introduces a comprehensive hardware platform based on multi-oxide RRAM crossbar arrays that achieves both precision for analog computing and stochasticity for intrinsic security and generative diversity. Leveraging 5-bit device-level analog programmability and reliable 4-bit array-level weight mapping, key primitives of memory-augmented neural networks (MANNs) are implemented, including convolutional encoding, locality-sensitive hashing, and RRAM-based CAM for few-shot learning with near-software accuracy. At the same time, inherent device variability is utilized as a high-entropy physical unclonable function for secure key generation and as a hardware-native randomness source that enhances the diversity and perceptual realism of StyleGAN3-generated biometric images. By integrating analog VMM, associative memory search, high-entropy randomness, and hardware-seeded generative models within the same crossbar fabric, this work demonstrates how “stochastic yet precise” memristor arrays can provide a unified foundation for edge-intelligent, secure, and data-generating systems. Controlled multi-bit programming, device-aware learning, and engineered randomness together enable a new class of memory-centric AI platforms that support inference, few-shot adaptation, PUF-based security, and high-fidelity generative AI.
Speaker(s): Sungjun Kim,
Virtual: https://events.vtools.ieee.org/m/534471