TY - GEN
T1 - Comparison of Invertible Architectures for High Speed Channel Design
AU - Bhatti, Osama Waqar
AU - Akinwande, Oluwaseyi
AU - Ambasana, Nikita
AU - Yang, Xianbo
AU - Paladhi, Pavel R.
AU - Becker, Wiren Dale
AU - Swaminathan, Madhavan
N1 - Funding Information:
IV. CONCLUSION In this work, we investigate the inverse design solutions for a high speed channel for 3 architectures: fully-connected neural networks, conditional generative adversarial networks and invertible neural networks. The quality of the posteriors generated by cGAN and INN are similar to actual posteriors. ACKNOWLEDGMENT This material is based upon work supported by the National Science Foundation under Grant No. CNS 16-24810 - Center for Advanced Electronics through Machine Learning (CAEML).
Publisher Copyright:
© 2021 IEEE.
PY - 2021
Y1 - 2021
N2 - In this paper, we perform a comparison of state-of-the-art inverse design strategies for a high-speed link. We take into account three machine learning architectures: (1) traditional fully-connected neural networks, (2) conditional generative adversarial networks and (3) invertible neural networks. The metrics for evaluation include quantitative mean-squared-errors on the test set as well as qualitative posterior distributions' similarity to the actual posterior distributions. We find that, on average, invertible neural networks have minimum mean-squared error for the input design tuple and their posterior shapes are in accordance with the actual distributions.
AB - In this paper, we perform a comparison of state-of-the-art inverse design strategies for a high-speed link. We take into account three machine learning architectures: (1) traditional fully-connected neural networks, (2) conditional generative adversarial networks and (3) invertible neural networks. The metrics for evaluation include quantitative mean-squared-errors on the test set as well as qualitative posterior distributions' similarity to the actual posterior distributions. We find that, on average, invertible neural networks have minimum mean-squared error for the input design tuple and their posterior shapes are in accordance with the actual distributions.
UR - http://www.scopus.com/inward/record.url?scp=85124143720&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85124143720&partnerID=8YFLogxK
U2 - 10.1109/EDAPS53774.2021.9657014
DO - 10.1109/EDAPS53774.2021.9657014
M3 - Conference contribution
AN - SCOPUS:85124143720
T3 - IEEE Electrical Design of Advanced Packaging and Systems Symposium
BT - 2021 IEEE Electrical Design of Advanced Packaging and Systems, EDAPS 2021
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2021 IEEE Electrical Design of Advanced Packaging and Systems, EDAPS 2021
Y2 - 13 December 2021 through 15 December 2021
ER -