dilated convolution

A convolutionconvolution, n.: An operation that slides a small grid of weights across an image or signal, multiplying and adding at each spot, to measure how strongly a pattern appears at every location. whose filterfilter, n.: In an image-recognition network, a small grid of learned numbers that acts as a pattern detector, for vertical edges, a patch of fur or a wheel, scanned across the whole picture. taps are spaced d positions apart (the dilation rate), so a kernel of size k spans d(k − 1) + 1 input positions without additional parameters or computation; d = 1 is ordinary convolution. As Fisher Yu and Vladlen Koltun emphasized, stacking layers whose dilation doubles (1, 2, 4, 8, …) widens the receptive fieldreceptive field, n.: The patch of the original input, such as a region of an image, that a particular neuron inside a network can see and respond to. exponentially with depth without any downsampling, which suits dense prediction such as semantic segmentationsemantic segmentation, n.: Coloring every pixel of an image by what it shows, such as road, sky, person or car, without telling one person or car apart from another. (the DeepLab family) and long-range audio modeling (WaveNetWaveNet, prop. n.: A DeepMind model that generates speech and other sounds one tiny audio sample at a time, producing voices far more natural than earlier computer speech.'s dilated causal convolutions). A known artifact is 'gridding', in which sparse sampling misses neighboring inputs; mixing dilation rates mitigates it.

Ten layers of dilated convolutions give each output a receptive field of over a thousand samples.