Implicit Bias of (Stochastic) Gradient Descent on Deep Linear Diagonal Networks
Modern machine learning models are often overparameterized, i.e., have more parameters than training data.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv