In the simulations in subsection 3.2, all linear networks have The backbone architecture consists of a stack of 12 Conv-LSTM modules, and each module contains 32 units (channels). The backbone architecture is illustrated in Figure 7. To demonstrate ARMA networks' applicability to image segmentation, we evaluate it on a challenging The network architecture is illustrated in Figure 15a. The experimental results are summarized in Table 5. Since image classifications tasks do not require convolu-tional layers to have large receptive fields, the learned autoregressive coefficients concentrate around 0, as shown in Figure 6.

Global information is essential for dense prediction problems, whose goal is to compute a discrete or continuous label for each pixel in the images. Traditional convolutional layers in neural networks, originally designed for image classification, are restrictive in these problems since their receptive fields are limited by the filter size. In this work, we propose autoregressive moving-average (ARMA) layer, a novel module in neural networks to allow explicit dependencies of output neurons, which significantly expands the receptive field with minimal extra parameters. We show experimentally that the effective receptive field of neural networks with ARMA layers expands as autoregressive coefficients become larger. In addition, we demonstrate that neural networks with ARMA layers substantially improve the performance of challenging pixel-level video prediction tasks as our model enlarges the effective receptive field.