Speaker
Description
Compositional concept generalization, the ability to understand novel combinations of previously learned concepts, remains a major challenge for multimodal AI systems. We present a multistage variational quantum learning framework that extends the Distributional Compositional Categorical (DisCoCat) model to vision-language reasoning for image–text pairs. The proposed approach first learns object representations from single object image–caption pairs and then freezes these representations in order to learning only relational concepts. This separation enables more effective compositional reasoning and reduces the number of trainable parameters.
We evaluate the framework on the CLEVR compositional benchmark using one hot, multi hot, CLIP angle, CLIP amplitude, and a novel Collage image encoding. Compared with classical CLIP and non multistage quantum baselines, the proposed multistage approach consistently improves the results while requiring only a few hundred trainable parameters instead of hundreds of millions (employed by CLIP). The results further show that amplitude encoding is more effective for learning individual objects, whereas angle encoding provides better relational generalization through its nonlinear feature mapping.
To assess hardware feasibility, we execute the trained circuits on IBM quantum emulators and IBM Marrakesh hardware, where quantum fidelity is estimated using the destructive SWAP test. The measured overlaps remain strongly correlated with the fidelities computed during training, demonstrating that the learnt similarity relationships can be preserved under realistic quantum execution. These results show that multistage variational quantum learning provides a parameter-efficient and hardware compatible framework for compositional concept generalization.
| Topic | Quantum variational circuits and their applications |
|---|