4.1 Convolutional Neural Network
61
x ij
d IJ
convolution
transposed convolution
Fig. 4.4 Relation between the convolution and the transposed convolution
Fig. 4.5 Handwritten characters generated by a DCGAN. Left: Checkerboard pattern is noticeable
at the beginning of learning. Right: Later in learning, it reached a level at which the pattern is not
noticeable
convolution. However, due to the nature of the definition, a unique pattern called
checkerboard artifact is often generated (see Figs. 4.5 left and 4.6).
In a deep neural network, if the training up to the input layer d I J has not
progressed—for example, if d I J becomes almost uniform—then the transposed
convolution will simply be just shifting and adding the filters in the calculation,
resulting in the checkerboard artifact. If the training of the input layer d I J progresses
well, the last d I J will not be uniform and the checkerboard pattern will not be
visible.
61
x ij
d IJ
convolution
transposed convolution
Fig. 4.4 Relation between the convolution and the transposed convolution
Fig. 4.5 Handwritten characters generated by a DCGAN. Left: Checkerboard pattern is noticeable
at the beginning of learning. Right: Later in learning, it reached a level at which the pattern is not
noticeable
convolution. However, due to the nature of the definition, a unique pattern called
checkerboard artifact is often generated (see Figs. 4.5 left and 4.6).
In a deep neural network, if the training up to the input layer d I J has not
progressed—for example, if d I J becomes almost uniform—then the transposed
convolution will simply be just shifting and adding the filters in the calculation,
resulting in the checkerboard artifact. If the training of the input layer d I J progresses
well, the last d I J will not be uniform and the checkerboard pattern will not be
visible.