Building Autoencoders in Keras mentions the following toy task for semi-supervised pre-training:
detail-context matching: being able to match high-resolution but small patches of pictures with low-resolution versions of the pictures they are extracted from
Where could I read more about this?