时 间：2025-04-03 15:30 - 16:30
报告人：王中雷 厦门大学 副教授
地 点：普陀校区理科大楼A1514
主持人：於州 华东师范大学 教授
摘 要：
For distributed learning, median-of-means estimators are popular for statistical inference against Byzantine attacks, but they suffer from inefficiency or even bias when the sample is asymmetrically distributed. We tackle this problem by proposing a bias-corrected Byzantine-robust estimator via Cornish-Fisher expansion, and it can be widely implemented in various gradient descent algorithms for state-of-the-art deep learning models. We rigorously demonstrate that the bias of the proposed estimator is much smaller than a vanilla median-of-means estimator and its variations under regularity conditions, and asymptotic properties of the proposed estimator are also established. The proposed estimator outperforms its alternatives numerically in terms of bias and variance under different synthetic setups, and performs the best when analyzing the CIFAR-10 dataset.
报告人简介：
王中雷，厦门大学王亚南经济研究院副教授，研究兴趣包括抽样调查以及人工智能在气象水文等科学的交叉应用。多篇成果发表于JRSS-B、JASA、Biometrika以及Nature Communications等国际知名期刊。