We introduce advancements over the Complex AutoEncoder (below) to create continuous and distributed object-centric representations of real-world images without supervision.

We present the Complex AutoEncoder – an object discovery approach that takes inspiration from neuroscience to implement distributed object-centric representations. After introducing complex-valued activations into a convolutional autoencoder, it learns to encode feature information in the activations’ magnitudes and object affiliation in their phase values.

We propose a new framework for causal discovery in which we learn to infer causal relations across samples with different underlying causal graphs but shared dynamics.

We highlight a fundamental problem in the loss used by previous approaches for contrastive object discovery: it cannot differentiate between a model that represents all the different objects in a scene and a model that (re-)represents the same object over and over. Our proposed solution: a set-based contrastive loss.