Multi-task learning holds the promise of less data, parameters, and time than training of separate models. We propose a method to automatically search over multi-task architectures while taking resource constraints into consideration. We propose a search space that compactly represents different parameter sharing strategies. This provides more effective coverage and sampling of the space of multi-task architectures. We also present a method for quick evaluation of different architectures by using feature distillation. Together these contributions allow us to quickly optimize for efficient multi-task models. We benchmark on Visual Decathlon, demonstrating that we can automatically search for and identify multi-task architectures that effectively make trade-offs between task resource requirements while achieving a high level of final performance.
@article{arxiv.1908.04339,
title = {Feature Partitioning for Efficient Multi-Task Architectures},
author = {Alejandro Newell and Lu Jiang and Chong Wang and Li-Jia Li and Jia Deng},
journal= {arXiv preprint arXiv:1908.04339},
year = {2019}
}