Mu-DisCoCat: A Variational Pipeline for Compositional Generalization on Quantum Processors

TL;DR

Mu-DisCoCat is a quantum machine learning framework that enables compositional generalization in multimodal settings by combining variational quantum circuits with compositional distributional semantics. It learns object representations from image-text pairs and uses them to infer relations in multi-object scenes, showing improved out-of-distribution relational accuracy compared to CLIP in simulations and maintaining fidelity on noisy quantum hardware.

Why It Matters

This work demonstrates a practical path for leveraging near-term quantum processors to solve complex compositional reasoning tasks, which are central to human-like AI. By bridging multimodal learning with quantum computation, it opens new avenues for scalable, generalizable AI systems on emerging quantum hardware.

Key Findings

• Mu-DisCoCat achieves compositional generalization by first learning stable object representations and then inferring relations between them in multimodal contexts.
• The model outperforms the CLIP baseline in out-of-distribution relational accuracy when evaluated using Uhlmann state fidelity in classical simulations.
• Hardware deployment on noisy quantum emulators and real devices shows strong correlation between simulated and executed results, confirming robustness to noise.