Our paper, “SCORE: Approximating Curvature Information under Self-Concordant Regularization” got accepted for publication in Computational Optimization and Applications (COAP). This paper proposes an approximation scheme for the Generalized Gauss-Newton (GGN) algorithm, as well as a new method for selecting adaptive step-sizes for the algorithm in the (strongly) convex settings. The approach has also been tested on neural network models.

Read full paper here.