Increased demands on computing performance have led to growing interest in new computing paradigms that could be more energy efficient or more suitable for solving certain classes of problems. Algorithms based on analog neuromorphic architectures show promise for tasks such as pattern recognition, but realizing physical device elements with both the required functionality and performance has been challenging. In my talk I will discuss our recent work demonstrating that by taking advantage of materials and phenomena from solid-state batteries, it is possible to realize three-terminal analog switching devices that possess all of the required attributes for a neuromorphic architecture: stable multi-level analog states, write linearity, low write noise, and extremely low switching voltage and energy. Furthermore, I will show how simulations of backpropagation (a well-known neuromorphic algorithm) using our experimentally measured device properties achieve the highest classification accuracy of any device to date. I will also discuss experimentally validated simulations of the device physics that establish the factors governing performance and serve to assess scalability and energy dissipation.
