Puzzling Success of Overparameterization: Lottery Tickets or Escape Dimensions?

- This is an EPFL research paper (a Swiss technical university) about why overparameterized neural networks — models with far more parameters than training data points — often generalize better, not worse, than smaller models. This contradicts classical statistical wisdom. - "Lottery ticket hypothesis" is a prominent theory: large networks work because they contain a small lucky "winning" subnetwork that does the actual learning. This paper argues that's not the full story. - "Escape dimensions" is the authors' alternative concept: extra parameters create new directions in the model's optimization landscape that help gradient descent escape bad local minima and find flatter, more generalizable solutions. - The work uses mathematical analysis on simplified models (two-layer ReLU networks) to compare these two explanations systematically. It speaks to a live debate in deep learning theory about why "bigger is better" actually works.