Learning Personalized Modular Network Guided by Structured Knowledge

摘要: The dominant deep learning approaches use a "one-size-fits-all" paradigm with the hope that underlying characteristics of diverse inputs can be captured via fixed structure. They also overlook importance explicitly modeling feature hierarchy. However, complex real-world tasks often require discovering reasoning paths for different to achieve satisfying predictions, especially challenging large-scale recognition label relations. In this paper, we treat structured commonsense knowledge (e.g. concept hierarchy) as guidance customizing more powerful and explainable network structures distinct inputs, leading dynamic individualized inference paths. Give an off-the-shelf large configuration, proposed Personalized Modular Network (PMN) is learned by selectively activating sequence modules where each them designated recognize particular levels knowledge. Learning semantic configurations activation align well regarded decision-making procedure, which solved new graph-based reinforcement algorithm. Experiments on three segmentation classification show our PMN superior performance reduced number while personalized module input.