TY - JOUR
T1 - Memristive recurrent neural network
AU - Tornez Xavier, Gerardo Marcos
AU - Gómez Castañeda, Felipe
AU - Flores Nava, Luis Martín
AU - Moreno Cadenas, José Antonio
N1 - Publisher Copyright:
© 2017
PY - 2018/1/17
Y1 - 2018/1/17
N2 - It is reported a continuous-time neural network in CMOS that uses memristors. These nanodevices are used to achieve some analog functions such as constant current sourcing, decaying term emulation, and resistive connection; all of them representing parameters of the neural network. The expected dynamics of this silicon circuit with these functional memristors is demonstrated via SPICE simulations based on 0.5 µm, n-well CMOS technology. The neural circuit is operative by finding the optimal solution of small-size combinatorial optimization problems, namely: “Assignment” and “Transportation”. It was chosen fast switching titanium dioxide memristors, which are modeled with nonlinear window functions and tunneling effect with the TEAM paradigm. This analog network belongs to an early recurrent model, which is electrically redesigned to take into account memristive arrays but keeping its original convergence properties. The behavioral and electrical analysis is done via Simulink-SPICE simulation. The outcome VLSI functional blocks combine both current and voltage to represent the variables in the recurrent model.
AB - It is reported a continuous-time neural network in CMOS that uses memristors. These nanodevices are used to achieve some analog functions such as constant current sourcing, decaying term emulation, and resistive connection; all of them representing parameters of the neural network. The expected dynamics of this silicon circuit with these functional memristors is demonstrated via SPICE simulations based on 0.5 µm, n-well CMOS technology. The neural circuit is operative by finding the optimal solution of small-size combinatorial optimization problems, namely: “Assignment” and “Transportation”. It was chosen fast switching titanium dioxide memristors, which are modeled with nonlinear window functions and tunneling effect with the TEAM paradigm. This analog network belongs to an early recurrent model, which is electrically redesigned to take into account memristive arrays but keeping its original convergence properties. The behavioral and electrical analysis is done via Simulink-SPICE simulation. The outcome VLSI functional blocks combine both current and voltage to represent the variables in the recurrent model.
KW - Analog VLSI design
KW - Continuous-time signal
KW - Hopfield
KW - Memristor
KW - Neural network
KW - Team model
UR - http://www.scopus.com/inward/record.url?scp=85028312618&partnerID=8YFLogxK
U2 - 10.1016/j.neucom.2017.08.008
DO - 10.1016/j.neucom.2017.08.008
M3 - Artículo
SN - 0925-2312
VL - 273
SP - 281
EP - 295
JO - Neurocomputing
JF - Neurocomputing
ER -