SLaK uses sparse, factorized kernels to scale convolutional networks to much larger receptive fields and evaluates the resulting models on image classification and dense prediction.