Weak Supervision for Object-Centric Visual Reasoning
A recent study presents a weakly supervised learning method aimed at grounding symbols within object-centric visual reasoning challenges. This technique merges a slot-based framework focused on object-centricity with a Variational Autoencoder (VAE) for self-supervised learning, rivaling concept guidance on latent dimensions to achieve human-readable grounding. The predictions generated are converted into symbolic background knowledge suitable for reasoning systems like Inductive Logic Programming (ILP) and Decision Trees. This innovative approach circumvents the expensive labeling typically needed in conventional two-stage neurosymbolic systems, all while preserving interpretability.
Key facts
- The paper is published on arXiv with ID 2605.08201.
- It introduces a weak supervision scheme for the perception stage of neurosymbolic systems.
- The approach uses a slot-based architecture for object-centricity.
- It employs a Variational Autoencoder (VAE) for self-supervision.
- Concept guidance is used on latent dimensions for human interpretable grounding.
- Predictions are translated into symbolic background knowledge for reasoning frameworks.
- Reasoning frameworks include Inductive Logic Programming (ILP) and Decision Trees.
- The method reduces the need for costly labels in perception output.
Entities
Institutions
- arXiv