Prof. Ohad Ben-Shahar

We describe a novel global optimization technique which utilizes global minima basins of attraction in order to quickly converge to a global minima. A key to the proposed method is the “steeper goes deeper” heuristic: coupling between magnitudes of gradients on different level sets of a basin of attraction and the depth of its minima. Local minima are avoided with a combination of local optimization and a heuristic-based leaping step. Gradient surfing performance is evaluated across a set of small-scale problems from the literature, and results are compared to those of 12 previously published methods. A practical six-dimensional non-convex image registration application is presented as well, where GS performance exceeds that of classic global optimization methods in both speed and accuracy. Additionally, we validate the optimization method by applying a new Gaussian mixture model benchmark for non-convex function. Finally, the “steeper goes deeper” heuristic is validated empirically on five different classes of non-convex functions using two different evaluation approaches. In all cases, steeper gradients are shown to lead to deeper optima with a high probability.