MEAM Seminar: “Nano-compatible Neuromorphic Computers: An All-passive Approach to Neural Networks”

This era of the internet of things is poised to experience explosive growth in the number of interconnected smart electronic devices. Machine learning algorithms, such as neural networks, are expected to play an indispensable role in analyzing the data gathered by these devices, and in many cases, will also facilitate informed device responses. However, the prospect of connecting a billion devices to the cloud and implementing large-scale neural networks remotely is infeasible for applications such as autonomous driving, implantable medical devices, and robotic platforms, that need the information to be processed speedily, at a small power budget. To deliver these requirements, it becomes necessary to equip devices with hardware that is optimized for neural network computations. For nanoscale implementations of such “neuromorphic” computers, it is important to develop simple designs of constituent circuits so that the required architectural complexity can be achieved within the nanofabrication constraints.

Current “neuromorphic” hardware designs typically utilize active circuits, comprising of three-terminal devices, to implement artificial neurons, an approach that is not suitable for compact nanoscale implementations. In this talk, we will show, how all-passive circuits for artificial neurons, comprising of two-terminal devices only, can instead address these drawbacks effectively. We will introduce simple circuit representations of artificial synapses and discuss how these enable improved memory-efficiencies vis – a – vis contemporary designs. By combining the passive neurons and synapses in network configurations, we will demonstrate how all-passive neuromorphic computers can perform complex pattern recognition tasks, such as identifying numerical digits from their images, with accuracies greater than 95%. The classification accuracies, power consumption and areal footprint of all – passive neuromorphic computers will be compared to those of the state – of – the – art technology. On the fabrication front, we will introduce simple methods for realizing re-programmable and once-programmable artificial synapses and delve into their operational attributes. Simple electrochemical deposition techniques for fabricating neurons will also be presented. The results of this work will promote new approaches to the design and fabrication of integrated nanoscale neuromorphic computers.