On Chip Customized Learning on Resistive Memory Technology for Secure Edge AI

This paper presents the first experimental demonstration of few-shot on-chip training on an in-memory computing resistive memory (ReRAM) platform. We use the Model-Agnostic Meta-Learning (MAML) algorithm to reduce training iterations and associated ReRAM conductance updates by orders of magnitude. Through co-optimization of device programming conditions and the algorithm, we achieve $>97 {\%}$ accuracy on the Omniglot dataset after just five training iterations (i.e., ReRAM programming operations) while improving device retention at 150°C.