Relative Contribution Mechanism: A Unified Paradigm for Disassembling Convolutional Neural Networks

Jiacong Hu, Jing Gao, Zunlei Feng, Lechao Cheng, Cheng Jin, Mingli Song

Abstract: With the tremendous development of CNNs, obtaining an available CNN classifier is more challenging due to the massive number of parameters and deep structure. Recently, an emerging model disassembling and assembling task (MDA-Task) has been proposed to obtain new models easily from the perspective of model reusing. However, the existing methods are usually slow or inaccurate. In this paper, we put forward a contribution paradigm for MDA-Task, which unifies existing model disassembling and assembling methods into a universal formulation. We first propose a relative contribution mechanism that the prediction results of the CNN classifier are decided by the larger contribution value. Then, the analysis and two discoveries of contribution allocation and aggregation procedures are given around the above mechanism. Based on the two discoveries, we introduce a contribution attribution based CNN disassembling technique composed of single-layer contribution attribution and backward accumulation attribution, which can effectively find the category-aware components in each layer and associated components in adjacent layers, respectively. In addition, a contribution rescaling based CNN assembling technique is devised for assembling the above disassembled category-aware components from different CNN classifiers, which can achieve comparable accuracy performance with original CNN classifiers. Experiments on five benchmark datasets with three mainstream CNN classifiers verify the effectiveness of the proposed contribution paradigm and demonstrate that the contribution attribution based CNN disassembling and assembling technique can achieve significant accuracy increases and faster speed than the existing methods.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)