Tailoring Resistive Switches for Applications in Artificial Intelligence and Beyond

Building artificial neuronal networks that mimic their biological counterparts is one of the remaining grand challenges in computing. Novel processors based on neural network architectures have recently been proposed by industry for commercial applications. However, they are based on existing digital technology, requiring dozens of transistor to implement the functionality of an analog synapse, thus limiting the system scalability on the long term. Over the past few years, I have been studying the emerging resistive switching technology (RRAM) – two terminal analog non-volatile electronic devices that have the potential to achieve high synaptic density comparable to biological neural circuits and with reasonable energy consumption. After a general overview of this emerging field, the talk will present recent results on three-dimensional RRAM crossbars, both standalone and monolithically integrated with CMOS circuitry, discussing issues related to manufacturability, reliability and yield. Several examples of using these devices for successfully prototyping promising applications in neuro-inspired computing, in-memory computing and security primitives will be highlighted. I will conclude by outlining my future research vision, focused on achieving dense energy-efficient hardware with applications in complex machine learning tasks and hybrid RRAM/biointerfaces.