Imperial College London


Faculty of EngineeringDepartment of Computing

Professor of Logic for Multiagent Systems



+44 (0)20 7594 8414a.lomuscio Website




504Huxley BuildingSouth Kensington Campus






BibTex format

author = {Hasani, R and Haerle, D and Baumgartner, C and Lomuscio, AR and Grosu, R},
publisher = {IEEE},
title = {Compositional neural-network modeling of complex analog circuits},
url = {},

RIS format (EndNote, RefMan)

AB - We introduce CompNN, a compositional method forthe construction of a neural-network (NN) capturing the dynamicbehavior of a complex analog multiple-input multiple-output(MIMO) system. CompNN first learns for each input/outputpair(i, j), a small-sized nonlinear auto-regressive neural networkwith exogenous input (NARX) representing the transfer-functionhij. The training dataset is generated by varying inputiof theMIMO, only. Then, for each outputj, the transfer functionshijare combined by a time-delayed neural network (TDNN) layer,fj.The training dataset forfjis generated by varying all MIMOinputs. The final output isf=(f1,...,fn). The NNs parame-ters are learned using Levenberg-Marquardt back-propagationalgorithm. We apply CompNN to learn an NN abstraction of aCMOS band-gap voltage-reference circuit (BGR). First, we learnthe NARX NNs corresponding to trimming, load-jump and line-jump responses of the circuit. Then, we recompose the outputsby training the second layer TDNN structure. We demonstratethe performance of our learned NN in the transient simulationof the BGR by reducing the simulation-time by a factor of 17compared to the transistor-level simulations. CompNN allows usto map particular parts of the NN to specific behavioral featuresof the BGR. To the best of our knowledge, CompNN is the firstmethod to learn the NN of an analog integrated circuit (MIMOsystem) in a compositional fashion.
AU - Hasani,R
AU - Haerle,D
AU - Baumgartner,C
AU - Lomuscio,AR
AU - Grosu,R
TI - Compositional neural-network modeling of complex analog circuits
UR -
ER -