A causal convolution lets the output at time \(t\) depend only on earlier inputs. 1-d: the dots are the inputs that reach the last output through a stack of causal dilated convolutions (WaveNet); the rows are the layers. 2-d: the output pixel is the black square and the image is read in raster order; blue pixels influence it, red pixels are in the causal region but can never be seen by a plain stack of masked convolutions (the blind spot), grey pixels are the part of the causal box covered by the two-stack design.
1-d: with kernel \(k\) and dilations \(d_1,\dots,d_L\) the last output sees \(1+(k-1)\sum_\ell d_\ell\) positions, a stack that doubles its dilation reaches back exponentially in the depth. 2-d: with half-width \(h=\lfloor k/2\rfloor\) and \(L\) layers the first layer must use mask A (the centre is excluded), the others mask B; the output then sees, in the row \(j\) above, only columns up to \(h\min(j,L)\) to the right, so a triangle of the causal box is missed.
code/ch22_autoregressive_models.py (receptive_field_1d, influence_2d, blind_spot, Figure x22.c).