Parallel Bayesian Active Learning using Dropout for Optimizing High-Speed Channel Equalization

This work realizes the parallelization of Bayesian Active Learning using Dropout (BAL-DO) and is successfully applied for optimizing equalization settings for high-speed channel receivers (RX). In this paper, parallel BAL-DO can achieve the largest horizontal eye (HEYE) opening and its corresponding equalization setting 12 times faster, on average, than the previously reported sequential BAL-DO. This corresponds to an average of 40–50 times faster than standard exhaustive time-domain simulations. Moreover, the HEYE prediction accuracy across the whole design space is close to 3 times better while using parallelization than sequential BAL-DO. With these outstanding results, parallel BAL-DO dramatically improves the efficiency for RX equalization optimization and greatly reduces the computing resources and time needed.