Groups
Category
In underdetermined linear systems (more variables than equations), gradient descent started at zero converges to the minimum Euclidean norm solution without any explicit regularizer.
Spectral analysis studies the distribution of eigenvalues and singular values of neural network weight matrices during training.