Fabrication And Application Of A Polymer Neuromorphic Circuitry Based On Polymer Memristive Devices And Polymer Transistors.

摘要

Neuromorphic engineering is a discipline that aims to address the shortcomings of today's serial computers, namely large power consumption, susceptibility to physical damage, as well as the need for explicit programming, by applying biologically-inspired principles to develop neural systems with applications such as machine learning and perception, autonomous robotics and generic artificial intelligence.;This doctoral dissertation presents work performed fabricating a previously developed type of polymer neuromorphic architecture, termed Polymer Neuromorphic Circuitry (PNC), inspired by the McCulloch-Pitts model of an artificial neuron. The major contribution of this dissertation is a development of processing techniques necessary to realize the Polymer Neuromorphic Circuitry, which required a development of individual polymer electronics elements, as well as customization of fabrication processes necessary for the realization of the circuitry on separate substrates as well as on a single substrate. This is the first demonstration of a fabrication of an entire neuron, and more importantly, a network of such neurons, that includes both the weighting functionality of a synapse and the somatic summing, all realized with polymer electronics technology.;Polymer electronics is a new branch of electronics that is based on conductive and semi-conductive polymers. These new elements hold a great advantage over the conventional, inorganic electronics in the form of physical flexibility, low cost and ease of fabrication, manufacturing compatibility with many substrate materials, as well as greater biological compatibility. These advantages were the primary motivation for the choice to fabricate all of the electrical components required to realize the PNC, namely polymer transistors, polymer memristive devices, and polymer resistors, with polymer electronics components.;The efficacy of this design is validated by demonstrating that the activation function of a single neuron approximates the sigmoidal function commonly employed by artificial neural networks. The utility of the neuromorphic circuitry is further corroborated by illustrating that a network of such neurons, and even a single neuron, are capable of performing linear classification for a real-life problem.