基于混合共享机制的多任务深度学习方法

Multi-task deep learning approach based on hybrid sharing mechanism

针对多任务学习的特征提取和任务区分难题,提出基于混合共享机制的多任务深度学习方法。给出硬参数共享网络混合训练方法和依据灵敏性分析的任务相似度分组策略;对组内、组间任务分别应用硬、软参数共享,给出混合共享网络及其相应训练方法;通过MNIST数据集上的实例研究与分析验证该方法的有效性。该方法充分发挥了硬、软参数共享机制的优点,较好刻画了任务的共享与私有特征,提升了多任务学习的性能。

Aiming at the problem of feature extraction and task classification in multi-task learning neural network, a multi-task deep learning method based on hybrid sharing mechanism was proposed. A hybrid training method of hard parameter sharing network and a task similarity grouping strategy based on sensitivity analysis were given. Hard parameter sharing was carried out for similar tasks in the same group, while soft parameter sharing for tasks with significant differences between groups, and the hybrid sharing network and its corresponding training method were provided. A case study and the comparative analysis on MNIST data set verified the effectiveness of the proposed learning approach. It gives full play to the advantages of hard and soft parameter sharing mechanism, describes the shared and private features of tasks more accurately, and improves the performance of multi-task learning model.