Title: Loss landscape of neural networks
Speaker: Dan Fletcher
Speaker Info: Northwestern
Brief Description: (remote talk)
Special Note:
Abstract:
A numerical analysis of the training dynamics of deep neural networks and comparison with the dynamics of spin glasses after a quench was discussed in Baity-Jesi et al. 2018 . This analysis suggested that barrier crossing does not take place during training of DNNs. I will present this analysis and also work of Freeman and Bruna 2017 which showed that for shallow networks, sub-level sets of the loss function are asymptotically path-connected.Date: Tuesday, April 28, 2020