This lecture: Spin-glass model (https://arxiv.org/abs/1412.0233, http://proceedings.mlr.press/v40/Choromanska15.pdf). Eliminating local minima (https://arxiv.org/abs/1901.00279).
Next lecture announcement:
Gradient descent (GD) almost surely converges to local minima for random initialization (https://arxiv.org/abs/1602.04915). Example of loss function for which GD converges exponentially slow (https://arxiv.org/abs/1705.10412). Noisy GD converges to local minima for any initialization (https://arxiv.org/abs/1503.02101) in polynomial time. Noisy GD converges to global minima (https://core.ac.uk/download/pdf/4380833.pdf).