PFGE: Parsimonious Fast Geometric Ensembling of DNNs

Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs

The loss functions of deep neural networks are complex and their geometric properties are not well understood. We show that the optima of these complex loss functions are in fact connected by a simple polygonal chain with only one bend, over which training and test accuracy are nearly constant. We introduce a training procedure to discover these high-accuracy pathways between modes. Inspired by...

Parsimonious Flooding in Geometric Random-Walks

We study the information spreading yielded by the (Parsimonious) 1-Flooding Protocol in geometric Mobile Ad-Hoc Networks. We consider n agents on a convex plane region of diameter D performing independent random walks with move radius ρ. At any time step, every active agent v informs every non-informed agent which is within distance R from v (R > 0 is the transmission radius). An agent is only ...

Fast Customization of Geometric Models

We present in this paper a technique for customization of geometric models of animals based on pictures. The object to be customized is divided into a set of hierarchically related local coordinate systems which allow for independent transformations of different body parts. A picture of the desired result is used as a visual aid to drive the customization process. We show the results of applyin...

Fast Pruning of Geometric Spanners

Let S be a set of points in R. Given a geometric spanner graph, G = (S, E), with constant dilation t, and a positive constant ε, we show how to construct a (1 + ε)-spanner of G with O(|S|) edges in time O(|E|+ |S| log |S|). Previous algorithms require a preliminary step in which the edges are sorted in non-decreasing order of their lengths and, thus, have running time Ω(|E| log |S|). We obtain ...

Network Evolution for DNNs