Nanoscale Single-electron Switching Arrays for Self-evolving Neuromorphic Networks

点击翻译按钮获取中文摘要

The goal of this project is to carry out a detailed multi-disciplinary study of single-electronlatching switches and of possible use of 2D arrays of such switches for hardwareimplementation of self-organizing (plastic) neuromorphic networks. Preliminary estimatesshow that such networks may provide unparalleled possibilities for complex informationprocessing. By these estimates, the networks may also have remarkable scaling properties:if implemented using a 10-nm technology, they may have density about 10 8 neurons percm 2 at manageable power dissipation below 100 W/cm 2 , and feature full learning cycle timeof the order of a few seconds. This scaling gives every hope that the networks will be able,after initial (largely unsupervised) learning, not only provide complex information processingincluding complex image recognition, but possibly reproduce biological evolution of thecerebral cortex at a time scale some 6 orders of magnitude shorter.The objective of the proposed project is to carry out a preliminary study of thisremarkable opportunity, addressing all its basic aspects at several structural levels. Inparticular, research will include the following components:A. Single-electron switch node design (D. Averin, K. Likharev, J. Wells).Detailed theoretical analysis and modeling (on two basic levels of single-electron transporttheory) of statics, dynamics, and statistics of the proposed single-electron latching switches.B. Low temperature prototyping (J. Lukens). Fabrication and experimentalstudy of Al/AlOx/Al prototypes of single-electron latching switches, with the goal to scalesingle-electron islands down to 100 nm and tunnel junctions to 10 nm, respectively, whichwould bring the reliable operation temperature up to about 10 K.C. Molecular single-electron device development (B. Brunschwig, J. Lukens,A. Mayr). Exploration of the opportunity to implement the basic component of the switches,the single-electron transistor, by chemical self-assembly of molecular components. Themolecular components will be deposited in solution on the prefabricated metallic wirestructures, and then characterized using a set of electrical, electrochemical, and time-resolvedlaser-spectrometry methods.D. Top level modeling and analysis (J. Barhen, M. Bender, K. Likharev).Large-scale computer simulation and a partial analytical study of the growth, dynamics, andself-adaptation of neuromorphic networks based on these switches.Hopefully, the project will achieve enough progress to justify a large-scale R&D effortin this exciting direction. In particular, a reliable evidence of self-organization of adaptiveneuromorphic networks during largely unsupervised learning would certainly be followed bythe first hardware implementations of sizable networks (possibly, after an initial stage ofpurely-CMOS-based prototyping using commercially available FPGA technology).The project will have a substantial educational component. Specifically (besidesparticipating in general educational Stony Brook initiatives), at least 4 FTE graduatestudents will be involved in the project each year, and some 20 undergraduate andgraduate students will take part in the project during its full 4-year period. At least onestudent will work in BNL and one in ORNL most of the time. Working in a multi-disciplinaryteam will allow these students to overcome inter-departmental barriers in their education.As another specific educational initiative, we plan to organize a Web-based undergraduatecourse on massively parallel supercomputing and neural networks, using the IBM SP3computer at Oak Ridge.Work on the inter-related aspects of this multi-disciplinary project will be constantlycoordinated by its P.I. (K. Likharev). In particular, regular meetings of all Stony Brook andBrookhaven participants of the team working on the project (including postdoctoralassociates and students), and annual meetings with Oak Ridge collaborators, are planned.