Transpose Convolution

When stride = 1, it’s overlaying kernels This is the ‘superposition’ interpretation of convolution it is the same as convolution with a flipped kernels

When stride > 1, it’s the same as conv(stretch(x))

See the notebook you emailed Julius

Filter the ‘mirrored’ copy of the signal

Gradient of Conv (makes forward and backwards same, natural way of upsampling)

the padding parameter is the same as truncating on both sides Last Reviewed: 1/3/25