Abstract
The inverted residual block is dominating architecture design for mobilenetworks recently. It changes the classic residual bottleneck by introducingtwo design rules: learning inverted residuals and using linear bottlenecks. Inthis paper, we rethink the necessity of such design changes and find it maybring risks of information loss and gradient confusion. We thus propose to flipthe structure and present a novel bottleneck design, called the sandglassblock, that performs identity mapping and spatial transformation at higherdimensions and thus alleviates information loss and gradient confusioneffectively. Extensive experiments demonstrate that, different from the commonbelief, such bottleneck structure is more beneficial than the inverted ones formobile networks. In ImageNet classification, by simply replacing the invertedresidual block with our sandglass block without increasing parameters andcomputation, the classification accuracy can be improved by more than 1.7% overMobileNetV2. On Pascal VOC 2007 test set, we observe that there is also 0.9%mAP improvement in object detection. We further verify the effectiveness of thesandglass block by adding it into the search space of neural architecturesearch method DARTS. With 25% parameter reduction, the classification accuracyis improved by 0.13% over previous DARTS models. Code can be found at:https://github.com/zhoudaquan/rethinking_bottleneck_design.