- Keywords: deep learning, optimization, non-convex optimization
- TL;DR: We investigate the effect of weight symmetry on the loss landscape of deep networks.
- Abstract: The loss landscapes of deep neural networks are poorly understood due to their high nonconvexity. Empirically, the local optima of these loss functions can be connected by a simple curve in model space, along which the loss remains fairly constant. Yet, current path finding algorithms do not consider the influence of symmetry in the loss surface caused by weight permutations of the networks corresponding to the minima. We propose a framework to investigate the effect of symmetry on the landscape connectivity by directly optimizing the weight permutations of the networks being connected. Through utilizing an existing neuron alignment technique, we derive an initialization for the weight permutations. Empirically, this initialization is critical for efficiently learning a simple, planar, low-loss curve between networks that successfully generalizes. Additionally, we introduce a proximal alternating minimization scheme to address if an optimal permutation can be learned, with some provable convergence guarantees. We find that the learned parameterized curve is still a low-loss curve after permuting the weights of the endpoint models, for a subset of permutations. We also show that there is small but steady performance gain in performance of the ensembles constructed from the learned curve, when considering weight space symmetry.
- Original Pdf: pdf