2026.10.9 9:00--10:00 报告一: Matrix Space Decomposition and Low-Rank Bilinear Representations

报告简介：This talk presents matrix space decomposition as a mathematical approach to compact representations in deep learning. Focusing on low-rank bilinear pooling, we discuss matrix factorization, subspace projections, and orthogonality constraints for reducing feature dimension while retaining discriminative information. We also examine the connection between principal component analysis of bilinear features and spectral clustering, highlighting how low-rank structure links efficient representation with discriminant learning.

2026.10.9 10:00--11:00 报告二: Decomposition of Nonlinear Invertible Maps for High-Dimensional Small-Sample Variational Inference

报告简介：This talk concerns the decomposition of nonlinear invertible maps for variational inference in high-dimensional, small-sample settings. We introduce bilateral normalizing flows, which use two coordinated lower-dimensional transformations to model data in sample and feature spaces. We discuss invertibility, exact likelihood evaluation, and structural constraints for reducing model complexity and stabilizing posterior approximation. Connections with low-rank modeling are also presented, highlighting the balance between expressive power and inference stability.

报告人简介：宋坤，哈尔滨工程大学助理研究员，博士。主要从事机器学习、矩阵空间建模、判别表示学习、计算机视觉与模式识别研究。2015年和2020年分别获西北工业大学硕士和博士学位，随后在穆罕默德·本·扎耶德人工智能大学（MBZUAI）从事机器学习研究。研究主要围绕矩阵空间分解与低秩结构建模、距离度量学习、二维矩阵分类及低秩双线性表示等。已发表学术论文十余篇，相关成果发表于IEEE TPAMI、IEEE TNNLS、Pattern Recognition等国际期刊及AAAI、IJCAI等国际会议。

报告地点：九游体育301