复旦大学-浙江师范大学“机器学习与视觉前沿论坛”系列报告三

报告题目：Structural Sparsity from Differential Inclusion to Deep Learning

报告专家：孙鑫伟（微软亚洲研究院）

报告时间：11月29日10:15-11:45

报告地点：腾讯会议号766-826-483（复旦浙师大MLV前沿论坛）

报告摘要：We consider recovering the signal with structural sparsity, i.e., the signal after a linear transformation is sparse. In this talk, we introduce a differential inclusion approach with variable splitting mechanism, which generates a regularization solution path. Equipped with the variable splitting, our method can alleviate the multicollinearity problem. Under this condition, we can prove the model selection consistency for this solution path. We further propose a data adaptive approach to determine a proper early stopping mechanism, towards controlling False-Discovery -Rate of selected features. Finally, we apply our method to explore the sparse structure of deep networks. Guided by differential inclusion, our method can forwardly learn the important weights/filters from sparse to dense, which can largely reduce the computational cost. Theoretically, we can prove the global convergence of this learning process.

邀请人： 郑忠龙