Sparse Relational Reasoning with Object-Centric Representations
This work addresses the interpretability-performance trade-off in relational AI models, but it is incremental as it builds on existing relational and object-centric methods.
The study examined how sparsity constraints affect the composability of soft-rules in relational neural architectures using object-centric representations, finding that increased sparsity improves performance and simplifies relations, but object-centric representations can fail if objects are not fully captured, highlighting trade-offs between interpretability and performance.
We investigate the composability of soft-rules learned by relational neural architectures when operating over object-centric (slot-based) representations, under a variety of sparsity-inducing constraints. We find that increasing sparsity, especially on features, improves the performance of some models and leads to simpler relations. Additionally, we observe that object-centric representations can be detrimental when not all objects are fully captured; a failure mode to which CNNs are less prone. These findings demonstrate the trade-offs between interpretability and performance, even for models designed to tackle relational tasks.