Flexible Multi-task Networks by Learning Parameter Allocation

Abstract

This paper proposes a novel learning method for multi-task applications. Multi-task neural networks can learn to transfer knowledge across different tasks by using parameter sharing. However, sharing parameters between unrelated tasks can hurt performance. To address this issue, we propose a framework to learn fine-grained patterns of parameter sharing. Assuming that the network is composed of several components across layers, our framework uses learned binary variables to allocate components to tasks in order to encourage more parameter sharing between related tasks, and discourage parameter sharing otherwise. The binary allocation variables are learned jointly with the model parameters by standard back-propagation thanks to the Gumbel-Softmax reparametrization method. When applied to the Omniglot benchmark, the proposed method achieves a 17% relative reduction of the error rate compared to state-of-the-art.

References in corpus (10)

• Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
• Neural Architecture Search with Reinforcement Learning
• DARTS: Differentiable Architecture Search
• PathNet: Evolution Channels Gradient Descent in Super Neural Networks
• Large-Scale Evolution of Image Classifiers
• Ask the GRU: Multi-Task Learning for Deep Text Recommendations
• Conditional Computation in Neural Networks for faster models
• SpotTune: Transfer Learning through Adaptive Fine-tuning
• Routing Networks and the Challenges of Modular and Compositional Computation
• Evolutionary-Neural Hybrid Agents for Architecture Search

Cited by in corpus (5)

• Learning to Branch for Multi-Task Learning
• DSelect-k: Differentiable Selection in the Mixture of Experts with Applications to Multi-Task Learning
• Routing Networks with Co-training for Continual Learning
• Deep Learning for Rheumatoid Arthritis: Joint Detection and Damage Scoring in X-rays
• Compositional Models: Multi-Task Learning and Knowledge Transfer with Modular Networks