Geometry-based Schr\"odinger Bridges for Trustworthy Multimodal Fusion
Researchers have developed a new method called Geometry-based Multimodal Fusion (GMF) to improve the trustworthiness of systems that combine data from multiple sources. Unlike existing methods that rely on a model's own confidence, GMF assesses data reliability by measuring the necessary correction in a latent space. This approach uses Diffusion Schrödinger Bridge transport to quantify how much adjustment is needed for input data, flagging unreliable inputs even when a model is confidently incorrect. Experiments show GMF significantly enhances robustness against sensor noise and conflicting data compared to traditional confidence-based baselines. AI
IMPACT Enhances the reliability of AI systems that process multiple data streams, crucial for real-world applications.