r/MachineLearning - [D] What is the SOTA explanation for why deep learning works? I understand backprop and gradient descent, but why should over-parametrized networks actually converge to anything useful?