Transfer Learning on Edge Using 14nm CMOS-compatible ReRAM Array and Analog In-memory Training Algorithm

This work presents a holistic approach to enabling energy-efficient on-chip Transfer Learning (TL) via Analog In-Memory Computing (AIMC) using 14nm CMOS-compatible ReRAM arrays. We develop an optimized ReRAM stack featuring H2 plasma-treated high-k (HfO2 or ZrO2) and in-vacuo processing, achieving reverse area scaling of forming voltage for co-integration with advanced-node CMOS technologies. To address non-ideal analog weight updates, we implement and evaluate the latest versions of Tiki-Taka training algorithms—TTv2, c-TTv2, and AGAD—capable of tolerating device asymmetry and variability. TL is demonstrated on hardware using compressed MNIST with on-chip training and extended via simulations to Vison Transformer (ViT)-based TL from CIFAR-10 to CIFAR-100. While analog-only models show sensitivity to weight transfer noise, hybrid analog-digital implementations maintain performance up to 20% noise. Using AGAD with optimized ReRAM devices, we achieve <1% accuracy degradation compared to digital baselines, validating AIMC-based TL as a viable path for low-power, on-chip training at the edge.