TOPIC PILLAR / COMPOSITIONALITY
Compositionality & Symbolic Reasoning
How do continuous representations in deep neural networks map to inspectable, discrete concepts that human operators can challenge and verify? Our research develops methods to extract structured knowledge trees and explanations at multiple levels of abstraction: quantising continuous representations into discrete codebooks, learning tree-structured visual semantics in hyperbolic space, and formulating model explanations as multi-agent argumentative debates (e.g. CVPRW, Discovery Science, AAMAS).
Articles & Research Explainers
Object-centric case-based reasoning via argumentation
SAA-CBR combines Slot Attention with Abstract Argumentation for Case-Based Reasoning to classify images through precedent-based dialectical reasoning.
Transparent visual reasoning via object-centric agent negotiation
OCEAN formulates visual reasoning as a multi-agent negotiation game where agents select and debate object explananda to reach transparent classifications.
Neuro-argumentative learning: uniting object slots with symbolic logic
Neural Argumentative Learning (NAL) marries Slot Attention with Assumption-Based Argumentation for transparent, rule-grounded image classification.
Causal composition in object-centric representations
Connecting object-centric learning with causal representation learning enables causally consistent scene generation and composition.
Free Argumentative Exchanges: multi-agent debates for visual explanations
Free Argumentative Exchanges (FAXs) enable agents to debate and internalise visual evidence using Bipolar Argumentation Frameworks.
Visual debates: explaining image classifiers through dialogue
Visual debates replace static saliency heatmaps with interactive argumentative dialogue between two agents.
Explaining an image classifier through a hierarchy of symbols
The method explains a trained image classifier using a learned hierarchy of discrete symbols.