METAL-OXIDE SYNAPSES WITH MULTIPLE-STATE-VARIABLE DYNAMICS
Abstract
Spiking neural networks (SNNs), the most biologically plausible neural networks, showpromise for building scalable, energy-efficient, artificial intelligence systems with fast learningperformance, thus addressing some of the challenges of currently dominating deep,learningneural networks (DNNs). Recent work showed the importance of integrating biologicalsynaptic plasticity mechanisms into SNNs,operation to achieve better functionality. Theproposal s overarching goal is to develop novel materials and devices for implementing,artificial synapses with plasticity mechanisms similar to biological counterparts, specificallymimicking neoHebbian plasticity.At th,e core of our proposal is mixed ionic-electronic materials with two (or more) types ofinteracting mobile ion species with drasticall,y different mobilities to emulate short- and longtermsynaptic memory. Mobile ionic species can be moved by the electric field (and p,ossiblyother types of external factors), which in turn result in a nonlinear change of electricalconductance that is suitable for mi,micking complex plasticity rules. For example, modulationof the concertation profile of faster ionic species could affect the mobili,ty of slower ionicspecies to emulate coupling between eligibility and synaptic weight state variables ofneoHebbian synapses.A partic,ular focus is ,ydrogen protons. In our preliminarywork, we have identified several candidate device structures for implementing neoHebbiansynapses., The most conservative (less challenging but sparser) approach involves two separate(slow and fast) titanium oxide memory devices th,at are integrated with conventional silicontransistors to provide the required coupling. A more advanced, much denser device is base,d ona single structure in which oxygen vacancy mobility is directly modulated by proton doping,e.g. by changing the vacancy charge s,tate.The proposed research involves coordinated theoretical and experimental tasks. During thefirst 12-month period of the project,,our main experimental effort will be to study hydrogenatednon-stoichiometric titanium dioxide (e.g., characterizing hydrogen mobilit,y and the impact ofhydrogen proton doping on the bulk/interface electrical conductance). The theoretical effortwill focus on numeric,al drift-diffusion modeling to aid experimental work and simulatepreliminary device structures.During the second 12-month period, we, plan to concentrate on developing techniques forcontrolling hydrogen proton and oxygen vacancy mobilities and their coupling, and d,evelopreproducible recipes for fabricating functional hydrogenated non-stochiometric titaniumdioxide films. In a parallel effort, we, will refine our numerical models using experimental dataand perform initial fabrication and testing of the most promising devices,,mainly focusing onsimple, planar structures.Finally, during the last, third 12-month period of the project, we will demonstrate the,mostpromising vertical device structure with directly coupled mobile species. The project willculminate in a prototype of a small (a,t least ten-device) network implementing a simplephoneme recognition task to demonstrate scalable learning functionality with neoHeb,biansynapses.If successful, our research will lead to very compact, at least 10x denser, neoHebbiansynapses with more robust and adj,ustable time scale switching dynamics compared to the priorwork. We believe that such a device will be crucial in developing the mos,t advanced SNNmodels with brain-like complexity and functionality, which are of primary importance for USmilitary superiority.Approv,ed for Public Release.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 07, 2022
- Source ID
- N000142212842
Entities
People
- Dmitri Strukov
Organizations
- Office of Naval Research
- United States Navy
- University of California, Santa Barbara