Recurrent Layer Aggregation using LSTM

Yu Qin, Jiajun Du, Xinyao Wang, Hongtao Lu

Abstract: Standard convolutional neural networks assemble multiple convolutional layers to extract high-level features. Recent efforts keep designing deeper and wider architectures. Even with skip connections applied to combine different layers, the useful low-level features are not effectively utilized. Some deep layer aggregation methods have been proposed to aggregate features of all levels, using simple linear combination or complex non-linear transformation. In this paper, we treat convolutional features as a sequence, and propose our Recurrent Aggregation of Convolutional Neural Network (CNN-RA). Our aggregation method splits a standard CNN into blocks and maps their feature matrices to a sequence of vectors of the same length. LSTM is employed to connect to the sequence and better fuse features across blocks. Our proposed CNN-RA can be directly appended to any standard CNN without any modifications. Experiments show remarkable improvements of CNN-RA over the original architectures across datasets.