LRN: Limitless Routing Networks for Effective Multi-task Learning

Ryan Wickman, Xiaofei Zhang, Weizi Li

Keywords: multi-task learning, MTL, reinforcement learning, machine learning, routing networks, modular networks

Abstract: Multi-task learning (MTL) is a field involved with learning multiple tasks simultaneously typically through using shared model parameters. The shared representation enables generalized parameters that are task invariant and assists in learning tasks with sparse data. However, the presence of unforeseen task interference can cause one task to improve at the detriment of another. A recent paradigm constructed to tackle these types of problems is the routing network, that builds neural network architectures from a set of modules conditioned on the input instance, task, and previous output of other modules. This approach has many constraints, so we propose the Limitless Routing Network (LRN) which removes the constraints through the usage of a transformer-based router and a reevaluation of the state and action space. We also provide a simple solution to the module collapse problem and display superior accuracy performance over several MTL benchmarks compared to the original routing network.

One-sentence Summary: A new multi-task learning framework that builds upon routing networks.