This research proposes the Perception-Reasoning Separation Network (PRS-Net), an architectural framework that introduces inductive biases to separate perception and reasoning in neural networks. The PRS-Net combines object-centric slot attention for perception with relational graph neural networks for reasoning, connected through a learned interface. The goal is to improve compositional generalization, interpretability, and out-of-distribution robustness compared to monolithic models.
Key findings
PRS-Net introduces explicit inductive biases to disentangle perception and reasoning.
The architecture combines object-centric slot attention with relational graph neural networks.
A learned interface bridges low-level visual features and abstract symbolic representations.
The model is expected to demonstrate improved compositional generalization and interpretability.
Limitations & open questions
The proposed architecture requires comprehensive validation on various benchmarks.
The effectiveness of the learned interface in different contexts needs further exploration.