To achieve near-zero training error in a classification problem, the layers of a feed-forward network have to disentangle the manifolds of data points with different labels, to facilitate the discrimination. However, excessive class separation can bring to overfitting since good generalisation requires learning invariant features, which involve some level of entanglement. We report on numerical experiments showing how the optimisation dynamics finds representations that balance these opposing tendencies with a non-monotonic trend. After a fast segregation phase, a slower rearrangement (conserved across data sets and architectures) increases the class entanglement.The training error at the inversion is stable under subsampling, and across network initialisations and optimisers, which characterises it as a property solely of the data structure and (very weakly) of the architecture. The inversion is the manifestation of tradeoffs elicited by well-defined and maximally stable elements of the training set, coined ``stragglers'', particularly influential for generalisation.
View on arXiv