Multimodal Deep Learning for Low-Resource Settings: A Vector Embedding Alignment Approach for Healthcare Applications

Objective Large-scale multi-modal deep learning models and datasets have revolutionized various domains such as healthcare, underscoring the critical role of computational power. However, in resource-constrained regions like Low and Middle-Income Countries (LMICs), GPU and data access is limited, leaving many dependent solely on CPUs. To address this, we advocate leveraging vector embeddings for flexible and efficient computational methodologies, aiming to democratize multimodal deep learning across diverse contexts. Background and Significance Our paper investigates the computational efficiency and effectiveness of leveraging vector embeddings, extracted from single-modal foundation models and multi-modal Vision-Language Models (VLM), for multimodal deep learning in low-resource environments, particularly in health-care applications. Additionally, we propose an easy but effective inference-time method to enhance performance by further aligning image-text embeddings. Materials and Methods By comparing these approaches with traditional multimodal deep learning methods, we assess their impact on computational efficiency and model performance using accuracy, F1-score, inference time, training time, and memory usage across 3 medical modalities such as BRSET (ophthalmology), HAM10000 (dermatology), and SatelliteBench (public health). Results Our findings indicate that embeddings reduce computational demands without compromising the model’s performance, and show that our embedding alignment method improves the performance of the models in medical tasks. Discussion This research contributes to sustainable AI practices by optimizing computational resources in resource-constrained environments. It highlights the potential of embedding-based approaches for efficient multimodal learning. Conclusion Vector embeddings democratize multimodal deep learning in LMICs, especially in healthcare. Our study showcases their effectiveness, enhancing AI adaptability in varied use cases.