Inversion dynamics of class manifolds in deep learning reveals tradeoffs underlying generalisation

Open in new window