The question of where a moving particle will come to rest on the landscape down which it is moving is a challenging one, especially if the terrain is rugged. In this paper we present experimental results of such a movement in two different contexts: spherical spin glasses and two-layer networks on the MNIST dataset. The unifying property of the different cases is that if the system is symmetric and complex enough, its landscape is trivial in the sense that random initial points and a chosen descent algorithm always lead to the same level of height on the terrain. This indicates the existence of energy barriers in large systems. We further test the existence of a barrier on a cost surface that theoretically has many zeroes. With this observation in mind we believe modifying the model to relocate this barrier is at least as important as finding a good descent algorithm on a fixed landscape.
View on arXiv