A1530
Title: Bayesian one-pass online learning for generalized linear models
Authors: Minwoo Chae - Pohang University of Science and Technology (Korea, South) [presenting]
Jeyong Lee - POSTECH (Korea, South)
Junhyeok Choi - Pohang University of Science and Technology (Korea, South)
Dongguen Kim - POSTECH (Korea, South)
Abstract: A Bayesian one-pass online learning algorithm is proposed for large-scale streaming data where observations are processed individually and exactly once. By utilizing a local quadratic approximation of the log-likelihood within a Gaussian variational framework, the method achieves computationally efficient, closed-form updates. An online analogue of the Bernstein-von Mises theorem is established, proving that the sequentially updated posterior is asymptotically equivalent to the full batch posterior and provides valid frequentist uncertainty quantification. The analysis and numerical experiments demonstrate that a critical warm-start phase ensures stability and allows the algorithm to match the efficiency of the batch MLE, consistently outperforming first-order methods like SGD.
