Publication
We study how neural networks compress uninformative input space in models where data lie in d dimensions, but the labels of which only vary within a linear manifold of dimension d(parallel to) < d. We show that for a one-hidden-layer network initialized with infinitesimal weights (i.e. in the feature learning regime) trained with gradient descent, the first layer of weights evolves to become nearly insensitive to the d(perpendicular to) = d - d(parallel to) uninformative directions. These are effectively compressed by a factor lambda similar to p, where p is the size of the training set. We quantify the benefit of such a compression on the test error epsilon. For large initialization of the weights (the lazy training regime), no compression occurs and for regular boundaries separating labels we find that epsilon similar to p(-beta), with beta(Lazy) = d/(3d - 2). Compression improves the learning curves so that beta(Feature) = (2d - 1)/(3d - 2) if d(parallel to) = 1 and beta(Feature) = (d + d(perpendicular to)/2)/(3d - 2) if d(parallel to) > 1. We test these predictions for a stripe model where boundaries are parallel interfaces (d(parallel to) = 1) as well as for a cylindrical boundary (d(parallel to) = 2). Next, we show that compression shapes the neural tangent kernel (NTK) evolution in time, so that its top eigenvectors become more informative and display a larger projection on the labels. Consequently, kernel learning with the frozen NTK at the end of training outperforms the initial NTK. We confirm these predictions both for a one-hidden-layer fully connected network trained on the stripe model and for a 16-layer convolutional neural network trained on the Modified National Institute of Standards and Technology database (MNIST), for which we also find beta(Feature) > beta(Lazy). The great similarities found in these two cases support the idea that compression is central to the training of MNIST, and puts forward kernel principal component analysis on the evolving NTK as a useful diagnostic of compression in deep networks.