Merging Heterogeneous Deep Models and Learning Retrieval Features
Academia Sinica Research Fellow Chu-Song Chen
Given different deep models that are well trained, we develop effective methods that can merge them into a single one. The merged model can handle the different tasks of the original ones simultaneously. Deep model merging has a great many potential of applications. Our approach can merge different deep models by removing the joint redundancy among them, so that a more compact model is obtainable for mobile or edge-computing usages. Besides, we also develop deep learning methods for constructing compact feature representations, so as to increase the retrieval efficiency in a large database.