Fitting and sharing multi-task learning

Multi-Task Learning is an effective method for learning cross-task knowledge. However, existing methods cannot fairly treat each task, their public parts are prone to continuously fit new tasks and decrease the performances of previous tasks. In this paper, we propose the Fitting-sharing Multi-Task Learning method to address this problem. In the Fitting step, a group of indicator parameters are trained to extract task-specific features and store them into an in-task template matrix. After all models converge, the indicators and templates are frozen to protect the learned knowledge. In the Sharing step, a group of connector parameters are trained to acquire information from other templates and to reason cross-task knowledge. Since the learning and sharing processes are separate, each model can acquire the learned knowledge from other tasks without affect them, and the imbalanced cross-task knowledge problem can be naturally avoided. Experimental results on public datasets illustrate that the proposed method can insistently improve the performance compared with existing methods.