Memristor crossbar circuits of unconventional computing for low-power IoT devices

In applications where power consumption is a critical factor, such as wearable devices, smart home automation, and remote monitoring systems, optimizing energy efficiency is crucial. This presentation examines the use of memristor crossbar circuits for the implementation of Internet of Things (IoT) devices capable of on-device learning and ondevice sensing. The first major contribution is a new analog circuit technique for equilibrium propagation (EP), which enhances on-device learning by calculating free-phase and nudge-phase solutions simultaneously, thus eliminating the need for intermediate storage. A simplified EP learning rule, based on fixed conductance changes, further improves the practicality of the system. The second contribution focuses on on-device sensing with a Dynamic Vision Sensor (DVS) camera integrated with the hybrid circuits. The system processes streaming events efficiently, converting DVS camera events into histogram data via input neurons, and performing low-power MAC calculations with memristor crossbars. The controller dynamically adjusts the MAC frequency and minimizes unnecessary power consumption. Simulations with POKER-DVS and MNIST-DVS datasets demonstrate high recognition rates ($\mathbf{9 7. 4 5 \%}$ and $\mathbf{9 1. 8 \%}$ respectively) while achieving up to 79% power savings, validating the potential of this approach for edge intelligence hardware.