5 problems

Deep neural networks are trained by gradient descent with small initialization; in this setting, the training dynamics and the notion of solution complexity are considered in the f…

For a neural network model, wider minima are local minima characterized by greater flatness or width in the loss landscape, and stochastic optimization refers to optimization metho…

A neural network estimator is obtained by optimizing an empirical loss over a class of deep neural networks. Optimizer-induced implicit regularization conjecture. It is conjectured…

Let $\mathbf X \in \mathbb{R}^{n \times d}$ have i.i.d. rows…

Nuclear norm implicit-regularization conjecture. Without the commutativity assumption, gradient descent should still find a solution of the reconstruction problem with minimum nucl…