Evenly Cascaded Convolutional Networks
share
In this paper we demonstrate that state-of-the-art convolutional neural networks can be constructed using a cascade algorithm for deep networks, inspired by the cascade algorithm in wavelet analysis. For each network layer the cascade algorithm creates two streams of features from the previous layer: one stream modulates the existing features producing low-level features, the other stream produces new features of a higher level. We evenly structure our network by resizing feature map dimensions by a consistent ratio. Our network produces humanly interpretable features maps, a result whose intuition can be understood in the context of scale-space theory. We demonstrate that our cascaded design facilitates the training process through providing easily trainable shortcuts. We report new state-of-the-art results for small networks - a consequence of our architecture's simple structure and direct training, without the need for additional treatment such as pruning or compression. Our 6-cascading-layer design with under 500k parameters achieves 95.24
READ FULL TEXT
