博士生讨论班2025[21]
时间：2025-10-28
阅读量：235次
- 演讲人： 高海亮
- 时间：2025年10月28日14:00
- 地点：浙江大学紫金港校区行政楼1417报告厅
报告文章：Conditional Matrix Flows for Gaussian Graphical Models (Marcello Massimo Negri & Fabricio Arend Torres & Volker Roth)
文章介绍：This paper introduces Conditional Matrix Flow (CMF), a unified variational framework for learning Gaussian Graphical Models. CMF integrates two components: (1) a matrix-variate normalizing flow, constrained by Cholesky decomposition to output symmetric positive definite precision matrices, and (2) a hypernetwork conditioning the flow on regularization strength $\lambda$ and $l_q$-norm parameter $q$. This design enables a single model to jointly learn a continuum of sparse models. By using simulated annealing, the framework unifies Bayesian posterior inference with frequentist MAP solution paths. CMF avoids MCMC costs and addresses $l_1$ over-shrinking by continuously exploring the entire $l_q$-norm family, including non-convex $q<1$.