Applications involving large-scale dictionary learning tasks motivate well online optimization algorithms for generally non-convex and non-smooth problems. In this big data context, the present paper develops an online learning framework by jointly leveraging the stochastic approximation paradigm with first-order acceleration schemes. The generally non-convex objective evaluated online at the resultant iterates enjoys quadratic rate of convergence. The generality of the novel approach is demonstrated in two online learning applications: (i) Online linear regression using the total least-squares approach; and, (ii) a semi-supervised dictionary learning approach to network-wide link load tracking and imputation of real data with missing entries. In both cases, numerical tests highlight the potential of the proposed online framework for big data network analytics.
