Contrastively-trained Structured World Models (C-SWMs) depart from traditional pixel-based reconstruction losses and use an energy-based hinge loss for learning object-centric world models.
By training on a large-scale repository of Visual Commonsense Graphs, VisualCOMET, a single stream vision-language transformer model, is able to generate inferences about past and present events by integrating information from the image with textual descriptions of the present event and location.
Demonstrates the potential of forward-prediction for solving PHYRE physical reasoning tasks by investigating various combinations of object and pixel-based forward-prediction and task-solution models.
The Transformer, a sequence transduction model that replaces recurrent layers and relies entirely on attention mechanisms, achieves new SotA on machine translation tasks while reducing training time significantly.
Hierarchical Relational Inference (HRI) learns hierarchical object representations and their relations directly from raw visual inputs, but is evaluated against limited baselines on simple datasets
Region Proposal Interaction Networks (RPIN) learn to reason about object trajectories in a latent region-proposal feature space, that captures object and contextual information.
Combines a compositional rendering network with a recurrent interaction network to learn dynamics in scenes with significant occlusion, but relies on ground-truth object positions and segmentations.
RELATE builds upon the interpretable, structured latent parameterization of BlockGAN by modeling the correlations between object parameters to generate realistic videos of dynamic scenes, using raw, unlabeled data.
Analysis of invariances in representations from contrastive self-supervised models reveals that they leverage aggressive cropping on object-centric datasets to improve occlusion invariance at the expense of viewpoint and category instance invariance.