Improving L-BFGS Initialization for Trust-Region Methods in Deep Learning

Deep learning algorithms often require solving a highly non-linear and nonconvex unconstrained optimization problem. Generally, methods for solving the optimization problems in machine learning and in deep learning specifically are restricted to the class of first-order algorithms, like stochastic gradient descent (SGD). The major drawback of the SGD methods is that they have the undesirable effect of not escaping saddle-points. Furthermore, these methods require exhaustive trial-and-error to fine-tune many learning parameters. Using the second-order curvature information to find the search direction can help with more robust convergence for the non-convex optimization problem. However, computing the Hessian matrix for the large-scale problems is not computationally practical. Alternatively, quasi-Newton methods construct an approximate of Hessian matrix to build a quadratic model of the objective function. Quasi-Newton methods, like SGD, require only first-order gradient information, but they can result in superlinear convergence, which makes them attractive alternatives. The limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) approach is one of the most popular quasi-Newton methods that construct positive-definite Hessian approximations. Since the true Hessian matrix is not necessarily positive definite, an extra initialization condition is required to be introduced when constructing the L-BFGS matrices to avoid false negative curvature information. In this paper, we propose various choices for initialization methods of the L-BFGS matrices within a trust-region framework. We provide empirical results on the classification task of the MNIST digits dataset to compare the performance of the trust-region algorithm with different L-BFGS initialization methods.