Routing Networks: Adaptive Selection of Non-linear Functions for\n Multi-Task Learning

Abstract

Multi-task learning (MTL) with neural networks leverages commonalities in\ntasks to improve performance, but often suffers from task interference which\nreduces the benefits of transfer. To address this issue we introduce the\nrouting network paradigm, a novel neural network and training algorithm. A\nrouting network is a kind of self-organizing neural network consisting of two\ncomponents: a router and a set of one or more function blocks. A function block\nmay be any neural network - for example a fully-connected or a convolutional\nlayer. Given an input the router makes a routing decision, choosing a function\nblock to apply and passing the output back to the router recursively,\nterminating when a fixed recursion depth is reached. In this way the routing\nnetwork dynamically composes different function blocks for each input. We\nemploy a collaborative multi-agent reinforcement learning (MARL) approach to\njointly train the router and function blocks. We evaluate our model against\ncross-stitch networks and shared-layer baselines on multi-task settings of the\nMNIST, mini-imagenet, and CIFAR-100 datasets. Our experiments demonstrate a\nsignificant improvement in accuracy, with sharper convergence. In addition,\nrouting networks have nearly constant per-task training cost while cross-stitch\nnetworks scale linearly with the number of tasks. On CIFAR-100 (20 tasks) we\nobtain cross-stitch performance levels with an 85% reduction in training time.\n