Why transposed convolution can be used to reconstruct activating features from the input images Cover image source A guide to convolution arithmetic for deep learning Vincent Dumoulin, Francesco Visin ArXiv In ZF Net, the transposed convolution is used to approximate the inverse of convolution, leading to the reconstruction of the activating features from the input image that activates a particular layer. However, strictly speaking, the transposed convolution is not really the inverse of convolution. Then why can it be used to do the reconstruction What is transposed convolution An ordinary convolution can be written in the form of Y C X . Then the corresponding transposed convolution is defined as X C T Y What yields the largest response The convolution Y C X and be transformed to the vectorized form y i c i T x where x is...