Stochastic positional embeddings improve masked image modeling

arXiv:2308.00566v2 Announce Type: replace-cross
Abstract: Masked Image Modeling (MIM) is a promising self-supervised learning approach that enables learning from unlabeled images. Despite its recent success, learning good representations through MIM remains challenging because it requires predicting the right semantic content in accurate locations. For example, given an incomplete picture of a dog, we can guess that there is a tail, but we cannot determine its exact location. In this work, we propose to incorporate location uncertainty into MIM by using …

abstract arxiv cs.ai cs.cv cs.lg dog embeddings example good image images locations modeling self-supervised learning semantic stochastic success supervised learning through type