Neural Teleportation

In this paper, we introduce neural teleportation, a simple operation one can use to initialize the weights of a neural network and gain faster convergence. Neural teleportation is the consequence of applying isomorphisms of quiver representations to neural networks. This process "teleports" a network to a new position in the weight space while leaving its input-to-output function unchanged. The concept of neural teleportation generalizes to any neural network architecture, activation function and task. We run several experiments that validate our hypothesis: teleporting a network at initialization speeds-up convergence. Finally, we discuss several mathematical and empirical findings concerning teleportation.