Neural networks and principal component analysis: Learning from examples without local minima

We consider the problem of learning from examples in layered linear feed-forward neural networks using optimization methods, such as back propagation, with respect to the usual quadratic error function E of the connection weights. Our main result is a complete description of the landscape attached to E in terms of principal component analysis. We show that E has a unique minimum corresponding to the projection onto the subspace generated by the first principal vectors of a covariance matrix associated with the training patterns. All the additional critical points of E are saddle points (corresponding to projections onto subspaces generated by higher order vectors). The auto-associative case is examined in detail. Extensions and implications for the learning algorithms are discussed.