Residual learning with skip connections permits training ultra-deep neural networks and obtains superb performance. Building in this direction, DenseNets proposed a dense connection structure where each layer is directly connected to all of its predecessors. The densely connected structure leads to better information flow and feature reuse. However, the overly dense skip connections also bring about the problems of potential risk of overfitting, parameter redundancy and large memory consumption. In this work, we analyze the feature aggregation patterns of ResNets and DenseNets under a uniform aggregation view framework. We show that both structures densely gather features from previous layers in the network but combine them in their respective ways: summation (ResNets) or concatenation (DenseNets). We compare the strengths and drawbacks of these two aggregation methods and analyze their potential effects on the networks' performance. Based on our analysis, we propose a new structure named SparseNets which achieves better performance with fewer parameters than DenseNets and ResNets.
View on arXiv