LAKE WALES, Fla. — Researchers at the University of Massachusetts have invented a diffusive-memristor that solves more accurately models the synapse of the brain by selectively forgetting little used information, while allowing vital information to be locked-in no matter how little it is accessed.
The problem with memristors today is that they perform deep learning too well. In real brains old knowledge that is no longer accessed gradually fades away, making room for the more important information that is accessed more frequently. This is no problem for Google-sized brains-in-the-cloud, but for self-sufficient mobile robots they need a memristor that emulates those in the human brain to keep from "filling up".
Nanoscale crossbar Pt/SiOxNy:Ag/Pt diffusive memristor (a) plus a scanning electron micrographs of the nanoscale crossbar junctions (b). URL: http://www.nature.com/nmat/journal/vaop/ncurrent/extref/nmat4756-s1.pdf
(Source: University of Massachusetts)
University of Massachusetts professor Joshua Yang, and colleagues, claim to have solved the problem with their diffusive memristor, which does everything that a regular memristor does, plus emulates the forgetting of old information no longer needed.
"Diffusive here refers to the conductance relaxation process in the diffusive memristor from high to low without electrical power, which is due to the diffusion of silver atoms (similar to calcium cations [Ca2+] in synapses) and responsible for some short-term plasticity (those can be forgotten)," Yang told EE Times.
In addition, the diffusive memristors can selectively lock-in important information in order to prevent its loss.
"Equally important, the diffusive memristor regulates the long-term plasticity in the 'nonvolatile' memristor connected to it, by providing the crucial synaptic dynamics (similar to those offered by Ca2+ dynamics in synapses). The long-term plasticity can be the memories you want to 'lock in'," Yang told EE Times.
For more information read Memristors with diffusive dynamics as synaptic emulators for neuromorphic computing
— R. Colin Johnson, Advanced Technology Editor, EE Times