Contrastive Training of Complex-Valued Autoencoders for Object Discovery

Neural Information Processing Systems 

Such slot-based approaches have several conceptual limitations.