SoftServe scales positive-definite quasi-Newton updates to deep networks

SoftServe replaces hard secant constraints with a soft variational objective that remains positive definite under negative curvature. Diagonal and Kronecker-factored variants target deep-learning scale without line searches.