A Sequential Multi-task Learning Neural Network with Metric-Based Knowledge Transfer
Simeng Yue, Seiichi Ozawa · 2012
In this paper, we propose a new sequential multitask pattern recognition model called Resource Allocating Network for Multi-Task Learning with Metric Learning (RAN-MTLML). RAN-MTLML has the following five functions: one-pass incremental learning, task-change detection, memory/retrieval of task knowledge, reorganization of classifier, and knowledge transfer. The knowledge transfer is actualized by transferring the metrics of all source tasks to a target task based on the task relatedness. Experimental results demonstrate the effectiveness of introducing the metric learning and the knowledge transfer on metric in the proposed RAN-MTLML.