Year: 2026
Type: article
Abstract: Multimodal learning aims to integrate complementary cues from different information sources to obtain more comprehensive and discriminative representations than those from a single modality. However, existing deep multimodal fusion methods often construct joint representations through feature concatenation, attention-based interactions or multi-stage fusion structures. Cross-modal conflicts are often implicitly compressed into high-dimensional features, which limits the interpretability of the fusion process. These designs may also introduce additional parameter overhead. To address these issues, this paper proposes a conflict-aware hybrid quantum residual network for multimodal fusion, namely the Hybrid Quantum Evidential Network (HQEN). HQEN extracts modality features through lightweight Dual-Modal Classical Encoders and uses an evidence-inspired shared-private decoupling mechanism to compress them into low-dimensional evidence descriptors with shared-private components. Before quantum encoding, conflict-aware phase mapping explicitly characterizes the conflict between the two modality descriptors. The resulting descriptors and conflict phase are encoded into different rotation channels of a parameterized quantum circuit. Rather than serving as an independent classifier, the quantum circuit acts as a quantum residual branch over the classical decision base, providing nonlinear correction in a low-dimensional evidence space. This design reduces the burden on a small-scale quantum circuit for complete multi-class prediction while keeping the model compact. Experimental results show that HQEN achieves competitive classification performance with a small parameter scale and a favorable accuracy--parameter trade-off. Ablation studies, conflict-stratified analysis and quantum residual output sensitivity analysis further support the roles of the evidence interface, conflict-aware phase and quantum residual branch. HQEN is trained and evaluated through classical quantum-circuit simulation, and the results support it as a compact hybrid quantum--classical modeling framework.
Source: Physica Scripta
Authors: Yihe Wang, Ri-Gui Zhou, Wenshan Xu, YaoChong Li
Institution: Shanghai Maritime University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
Awards: G3581250527, G5329798709, G7088808528
OpenAlex ID: W7235582401
