Quantum Feature Augmentation Boosts Multimodal Classification.

Most multimodal learning approaches focus on aligning and fusing heterogeneous data representations, but less attention has been paid to enhancing these features *after* fusion. This research proposes Parallel Quantum Feature Augmentation (PQFA), a novel hybrid quantum-classical framework designed to address this gap. PQFA takes fused multimodal features, derived from text and image encoders processed through cross-attention and adaptive gated fusion, and feeds them into parallel shallow variational quantum circuits. The quantum circuits' measurement readouts are then concatenated with the classical representation for final prediction. Evaluations on datasets like MM-IMDb and N24News show that PQFA consistently outperforms classical baselines, including the fusion backbone without augmentation and a width-matched MLP augmentation, while using significantly fewer parameters. Crucially, PQFA also demonstrates improved robustness in scenarios with missing modalities, particularly when critical textual input is degraded. This indicates that quantum augmentation offers a parameter-efficient and effective strategy for post-fusion enhancement in hybrid multimodal learning.