TY - GEN
T1 - Bayesian Active Learning for Uncertainty Quantification of High Speed Channel Signaling
AU - Torun, Hakki M.
AU - Hejase, Jose A.
AU - Tang, Junyan
AU - Beckert, Wiren D.
AU - Swaminathan, Madhavan
N1 - Funding Information:
shows the model confidently selects CPU RX package corner impedance as the dominant variable that affects HEYE since lower confidence bound (LCB) of its weight is higher than upper confidence bound (UCB) of the others. However, high uncertainty associated with the prediction of its weight indicates more data is required to calculate its exact weight. V. CONCLUSION In this work, we have introduced a new algorithm, Bayesian Active Learning using Dropout (BAL-DO), that simultaneously creates an accurate predictive probabilistic model and optimizes the function to perform a complete uncertainty quantification of an industrial high-speed channel. By jointly performing learning and optimization using BAL-DO, we have shown that both optimization performance and predictive accuracy of the model can be increased compared to BAL and d-TSBO, which focus on only learning and optimization, respectively. Although we have applied BAL-DO to high-speed channels, it has potential to be applied to other systems that can be represented as a black-box. ACKNOWLEDGEMENTS This work was partly supported by NSF I/UCRC Center for Advanced Electronics through Machine Learning (CAEML). The authors would like to thank Dr. Zhaoqing Chen and Mr. Troy Beukama from IBM for PCB trace corner modeling and for providing HSSCDR support.
Publisher Copyright:
© 2018 IEEE.
PY - 2018/11/13
Y1 - 2018/11/13
N2 - Increasing data rates in server high-speed communication busses makes their performance more susceptible to uncertainties in manufacturing processes. As a result, it is essential to understand channel design limitations and performance under tolerances to ensure a robust system. Predicting channel performance under tolerances can become very straining in time and computational resources. To address this, we propose a new active learning based algorithm that starts with no training data to simultaneously derive an accurate predictive model while finding the worst case scenario to ensure channel compliance in reduced CPU time compared to conventional methods.
AB - Increasing data rates in server high-speed communication busses makes their performance more susceptible to uncertainties in manufacturing processes. As a result, it is essential to understand channel design limitations and performance under tolerances to ensure a robust system. Predicting channel performance under tolerances can become very straining in time and computational resources. To address this, we propose a new active learning based algorithm that starts with no training data to simultaneously derive an accurate predictive model while finding the worst case scenario to ensure channel compliance in reduced CPU time compared to conventional methods.
UR - http://www.scopus.com/inward/record.url?scp=85059019156&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85059019156&partnerID=8YFLogxK
U2 - 10.1109/EPEPS.2018.8534251
DO - 10.1109/EPEPS.2018.8534251
M3 - Conference contribution
AN - SCOPUS:85059019156
T3 - EPEPS 2018 - IEEE 27th Conference on Electrical Performance of Electronic Packaging and Systems
SP - 311
EP - 313
BT - EPEPS 2018 - IEEE 27th Conference on Electrical Performance of Electronic Packaging and Systems
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 27th IEEE Conference on Electrical Performance of Electronic Packaging and Systems, EPEPS 2018
Y2 - 14 October 2018 through 17 October 2018
ER -