Learning theory for deep nets
主 题: Learning theory for deep nets
报告人: 林绍波 (温州大学)
时 间: 2016-12-16 10：00-11：00
地 点: 理科一号楼 1114
Deep learning has attracted avid research activities in the past few years. Compared with comprehensive application studies, the theoretical verifications lag heavily behind. This talk aims at developing a learning theory for deep learning to illustrate the power of deep nets. We construct a deep net containing pre-training stage, learning stage and fine-tuning stage to embody the three features of deep learning: multi-layered neural networks, large-scale algorithms and fine-tuning. Our constructed deep net is proved to attain the optimal learning rate when the ambient space is a lower dimensional manifold. This optimal learning rate is better than the existing results for shallow nets and therefore, shows the outperformance of deep nets.
