TY - GEN
T1 - Power-Efficient Deep Neural Networks with Noisy Memristor Implementation
AU - Dupraz, Elsa
AU - Varshney, Lav R.
AU - Leduc-Primeau, Francois
N1 - Publisher Copyright:
© 2021 IEEE.
PY - 2021
Y1 - 2021
N2 - This paper considers Deep Neural Network (DNN) linear-nonlinear computations implemented on memristor cross-bar substrates. To address the case where true memristor conductance values may differ from their target values, it introduces a theoretical framework that characterizes the effect of conductance value variations on the final inference computation. With only second-order moment assumptions, theoretical results on tracking the mean, variance, and covariance of the layer-by-layer noisy computations are given. By allowing the possibility of amplifying certain signals within the DNN, power consumption is characterized and then optimized via KKT conditions. Simulation results verify the accuracy of the proposed analysis and demonstrate the significant power efficiency gains that are possible via optimization for a target mean squared error.
AB - This paper considers Deep Neural Network (DNN) linear-nonlinear computations implemented on memristor cross-bar substrates. To address the case where true memristor conductance values may differ from their target values, it introduces a theoretical framework that characterizes the effect of conductance value variations on the final inference computation. With only second-order moment assumptions, theoretical results on tracking the mean, variance, and covariance of the layer-by-layer noisy computations are given. By allowing the possibility of amplifying certain signals within the DNN, power consumption is characterized and then optimized via KKT conditions. Simulation results verify the accuracy of the proposed analysis and demonstrate the significant power efficiency gains that are possible via optimization for a target mean squared error.
UR - http://www.scopus.com/inward/record.url?scp=85123434351&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85123434351&partnerID=8YFLogxK
U2 - 10.1109/ITW48936.2021.9611431
DO - 10.1109/ITW48936.2021.9611431
M3 - Conference contribution
AN - SCOPUS:85123434351
T3 - 2021 IEEE Information Theory Workshop, ITW 2021 - Proceedings
BT - 2021 IEEE Information Theory Workshop, ITW 2021 - Proceedings
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2021 IEEE Information Theory Workshop, ITW 2021
Y2 - 17 October 2021 through 21 October 2021
ER -