DeepLearningforFixedModel.PDFVIP

Deep Learning for Fixed Model Reuse? Yang Yang, De-Chuan Zhan, Ying Fan, Yuan Jiang and Zhi-Hua Zhou National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, 210023, China Collaborative Innovation Center of Novel Software Technology and Industrialization, Nanjing, 210023, China {yangy, zhandc, fany, jiangy, zhouzh}@ Abstract other tasks, rather than building a model from scratch. This offers a great potential to reduce the required amount of Model reuse attempts to construct a model by utilizing ex- training examples and training time cost, because the ex- isting available models, mostly trained for other tasks, rather ploitation of existing models may help set a good basis for than building a model from scratch. It is helpful to reduce the time cost, data amount, and expertise required. Deep learning the training of a new model. An example has been shown has achieved great success in various tasks involving images, in (Li, Tsang, and Zhou 2013), where a model aims to op- voices and videos. There are several studies have the sense of timizing AUC can be obtained with much less effort by model reuse, by trying to reuse pre-trained deep networks ar- reusing a model optimizing accuracy. Note that model reuse chitectures or deep model features to train a new deep model. also reduces the requirement of expertise in training the They, however, neglect the fact that there are many other ?xed models, because the user can start from a good model gen- models or features available. In this paper, we propose a more erated by experts, and thus, an expert-level new model can thorough model reuse scheme, FMR (Fixed Model Reuse). be obtained though the user him/herself is not an expert. FMR utilizes the learning power of deep models to implic