Multimodal vision‑language systems typically fuse image and text embeddings using classical operators such as concatenation, attention, bilinear pooling, or tensor interactions. QEMFN introduces a hybrid quantum‑classical framework that injects trainable entanglement as a structured inductive bias for fusion. Pre‑trained visual and textual features are first projected into compact latent spaces and encoded as angle‑parameterized quantum states. Intra‑modal and paired cross‑modal entangling circuits then act on these states, and measurement yields fused representations for retrieval.
With matched parameter budgets and frozen CLIP backbones, QEMFN outperforms classical baselines—including multilayer perceptron, tensor fusion, FiLM, cross‑attention, compact transformer, and a dequantized paired‑topology analogue—on COCO‑5k and Flickr30k. An ablation suite isolates the quantum module’s contribution from surrounding classical projections, while quantum‑centric analyses report Meyer‑Wallach entanglement capability, expressibility, gradient variance against barren‑plateau bounds, and entropy‑performance correlation under controlled training progress, accompanied by an intervention study on the entangling component.
QEMFN is executed using shot‑based estimation on a noise‑modeled simulated backend and a real superconducting device, employing zero‑noise extrapolation. This work does not claim quantum computational advantage; its contribution lies in the framework and a controlled empirical and quantum‑centric evaluation that positions trainable entanglement as an interpretable, hardware‑executable fusion mechanism at scales accessible on contemporary devices.
Review