Shared Residual Mapping Model for Efficient Ensemble

The multi-input multi-output (MIMO) model is an ensemble model with a single network for image classification. This model is optimized to create independent subnetworks in the backbone network and uses them for the ensemble of image classification. The problem with the MIMO model is that it can hold only a few subnetworks to keep the entire generalization. To overcome this problem, this study introduces the shared residual mapping (ShaRM) model. The ShaRM model consists of a single network with multiple information paths to increase the network capacity depending on the ensemble size. Because the ShaRM model shares residual mapping among input images, the ShaRM model effectively reduces the computational cost and becomes as efficient as the MIMO model. We optimize the ShaRM model on the CIFAR-100 dataset by changing the ensemble sizes and experimentally show that the ShaRM model is more robust against the ensemble size than the MIMO model. We also show that when the backbone network is relatively small, the ShaRM model achieves better accuracy than the MIMO model.