- This event has passed.
MEAM Thesis Defense: “Inkjet Printed Neuromorphic Inference Circuits with Memristor-Based Neuron Network”
April 30 at 10:30 AM - 11:30 AM
The deployment of machine learning inference algorithms on Internet of Things (IoT) devices remains challenging. Despite the low fabrication cost, flexibility, and low power consumption of the printed electronics for IoT applications, there are not many demonstrations of printed electronics solving neural network tasks, mainly due to the poor electrical performance, low device yield, and large footprint of printed thin-film transistors. In this study, we design and fabricate an inkjet-printed all-passive neuromorphic circuit based on printable memristor neuron architecture. The fully inkjet printed memristor is based on Ag/Poly(3,4-ethylenedioxythiophene): poly(styrenesulfonate) (PEDOT:PSS)/Ag. It shows bipolar resistance switching with low switching voltage (0.1~0.2V), good stability (9 days), high cyclability (200 cycles), and a large ON-to-OFF resistance margin (~40). A write-once-read-many times (WORM) memory is also observed in the proposed memristor when a large voltage (~3V) is applied. To present the neuromorphic computing capability, we fabricated the circuit that can realize the XOR classification problem with 100% accuracy, 0.7 cm2 size, and power consumption of 0.68 mW, by integrating the inkjet printed memristors with inkjet printed silver resistors. After establishing the device model of the inkjet printed memristors, we demonstrate via SPICE circuit simulations how a shallow network of memristor-resistors-based passive neuron networks can solve the IRIS classification problem with 95% accuracy and 2 mW power consumption. By providing the fabrication and simulation of the proposed memristor-based shallow multilayer perceptron, this thesis paves the way for low-power and low-cost neuromorphic inference devices for future applications of IoT.
MSE Candidate, Department of Mechanical Engineering and Applied Mechanics, University of Pennsylvania
Advisor: Mark G. Allen