Masked Autoencoder

A class of self-supervised reconstruction models in which a portion of the input is randomly masked during training and the model is required to reconstruct the masked content from the remaining visible portion, under the core assumption that a model able to accurately reconstruct a data distribution has internalized its regularities.