TOPIC PILLAR / OBJECT-CENTRIC LEARNING
Object-Centric Learning (OCL) & Disentanglement
Standard deep neural networks tend to exploit spurious visual correlations and contextual shortcuts instead of learning robust physical concepts. Object-Centric Learning (OCL) decomposes continuous perceptual scenes into structured, identifiable slots representing individual entities, attributes, and relationships. Our research addresses key theoretical bottlenecks: identifiability guarantees under severe spatial and occlusion ambiguities, continuous-discrete interfaces, and integrating object slots with symbolic argumentation frameworks (e.g. ICML, IJCAI, and NeSY).
Articles & Research Explainers
Resolving spatial ambiguities in object-centric representations
View-Invariant Slot Attention (VISA) resolves partial occlusions and multi-view ambiguities with theoretical identifiability guarantees.
Towards human-like visual reasoning: an object-centric foundation
A comprehensive synthesis of object-centric representation learning, theoretical identifiability, and causal reasoning.
Probabilistic Slot Attention: learning identifiable object representations
Probabilistic Slot Attention (PSA) introduces a generative formulation over object slots, proving identifiability without supervised masks.
Grounded Object-Centric Learning: giving slots a reusable vocabulary
Conditional Slot Attention learns a Grounded Slot Dictionary for reusable object representations.
LiTS: why a good liver outline can still miss a tumour
LiTS evaluates liver and tumour segmentation on a shared collection of abdominal CT scans.