- Learning composition: Improving object-centric learning by injecting compositionality
Authors: Jeong Ui, Yoo Jae-hoon, Ahn Sung-jin, Hong Seung-hoon
Abstract: Learning compository representations is a key aspect of object-centered learning because it enables flexible and systematic generalization and supports complex visual reasoning. However, most existing approaches rely on an auto-encoding objective, whereas compositivity is implicitly imposed by encoder architecture or algorithmic biases. This mismatch between the auto-encoding objective and learning compositivity often leads to failure to capture meaningful object representations. In this work, we propose a new objective that explicitly promotes compositivity of representations. Building on existing object-centered learning frameworks (e.g., slot attention), our method incorporates the additional constraint that any mixture of object representations from two images is valid by maximizing the likelihood on the synthetic data. We demonstrate that incorporating our objective into existing frameworks consistently improves objective-centered learning and is more robust to architecture choices.
2. Entity-centric Reinforcement Learning for Object Manipulation from Pixels
Authors: Dan Haramaty, Tal Daniel, Aviv Tamar
Abstract: Object manipulation is a hallmark of human intelligence and a key task in fields such as robotics. In principle, reinforcement learning (RL) offers a general approach to learning object manipulation. However, in practice, due to the curse of dimensionality, domains with more than a few objects are challenging for RL agents, especially when learning from raw image observations. In this work, we propose a structured approach for visual RL that is suitable for representing multiple objects and their interactions, and use it to learn goal-conditional manipulation of multiple objects. Key to our method is the ability to handle goals with dependencies between objects (e.g., moving objects in a specific order). Furthermore, based on theoretical results on constructive generalization, we link our architecture to the generalization capabilities of trained agents and demonstrate an agent that generalizes to similar tasks of more than 10 objects while learning on 3 objects. Videos and code are available on the project website: https://sites.google.com/view/entity-centric-r
