An adaptive multi-teacher-student knowledge distillation framework for scalable cloud–edge multimodal learning

Cloud–edge collaborative intelligence has become a key paradigm for deploying multimodal artificial intelligence (AI) applications in resource-constrained environments, enabling computationally intensive cloud models to support lightweight edge devices for real-time decision-making. However, existing multimodal knowledge distillation approaches often rely on static teacher–student transfer mechanisms, manually predefined distillation weights, and limited cross-modal adaptation capabilities, resulting in suboptimal knowledge transfer and reduced deployment efficiency on heterogeneous edge platforms. This study proposes the Adaptive Multi-Teacher Contrastive Distillation Network (AMTCD-Net), a novel cloud–edge multimodal intelligence framework designed for efficient knowledge transfer across image, text, and audio modalities. The proposed architecture integrates three key components: (i) a hypernetwork-driven adaptive weighting module that dynamically optimizes modality-specific distillation coefficients during training, (ii) a contrastive semantic alignment mechanism that preserves cross-modal representation consistency between large cloud-based teachers-student framework, and (iii) a dynamic teacher reliability selector that prioritizes the most informative teacher signals based on modality characteristics and deployment conditions. The cloud-based teacher models consist of Contrastive Language–Image Pretraining (CLIP) and Flamingo, while the lightweight edge student models comprise MobileNetV3 (Mobile Network Version 3) and Tiny Bidirectional Encoder Representations from Transformers (TinyBERT). The framework was evaluated using the publicly available MS COCO, VQA v2, and AudioSet benchmarks and further validated on NVIDIA Jetson Nano and Raspberry Pi 5 edge platforms to assess real-world deployment feasibility. Experimental results demonstrate that AMTCD-Net consistently outperforms conventional knowledge distillation, adaptive multi-teacher, and recent cross-modal learning baselines, achieving 88.9% image–text retrieval accuracy, 82.6% visual question answering accuracy, and 90.8% audio–visual classification accuracy. The proposed framework reduces edge-model computational cost by 68.3% while maintaining high predictive performance. Deployment experiments confirm practical suitability for real-time edge intelligence.