Spectral Analysis of Kernel and Neural Embeddings: Optimization and Generalization

We extend the recent results of (Arora et al. 2019). by spectral analysis of the representations corresponding to the kernel and neural embeddings. They showed that in a simple single-layer network, the alignment of the labels to the eigenvectors of the corresponding Gram matrix determines both the convergence of the optimization during training as well as the generalization properties. We generalize their result to the kernel and neural representations and show these extensions improve both optimization and generalization of the basic setup studied in (Arora et al. 2019). In particular, we first extend the setup with the Gaussian kernel and the approximations by random Fourier features as well as with the embeddings produced by two-layer networks trained on different tasks. We then study the use of more sophisticated kernels and embeddings, those designed optimally for deep neural networks and those developed for the classification task of interest given the data and the training labels, independent of any specific classification model.