DanielFilan comments on Visible loss landscape basins don’t correspond to distinct algorithms

[DanielFilan]

Update: I currently think that Nguyen (2019) proves the claim, but it actually requires a layer to have two hidden neurons per training example.