Tags:area reduction, convex polygon, ReLU network and shape approximation
Abstract:
Motivated by the study of shallow rectifier neural networks, we propose an algorithm that approximates a polygon in the plane by an outer polygon with fewer vertices. The algorithm minimizes the area locally in an iterative manner, and hence is guaranteed to find a local minimum and not necessarily the global optimum. Nevertheless, experiments indicate that in practice it yields a reasonably good approximation.