Quantifying the Uncertainty of Foundation Models with Singular Value Ensembles
Two new research papers explore methods for improving uncertainty quantification in foundation models. The first paper introduces Singular Value Ensemble (SVE), a parameter-efficient technique that modulates singular values of weight matrices to create diverse model ensembles, significantly reducing computational cost while maintaining accuracy and improving calibration. The second paper empirically compares tabular foundation models, specifically TabPFN, against Gaussian processes, revealing that while TabPFN excels in complex, data-rich scenarios, Gaussian processes offer superior performance and uncertainty quantification in data-scarce environments, especially when their kernel aligns well with the underlying function. AI
IMPACT Advances in uncertainty quantification are crucial for deploying foundation models in safety-critical applications, potentially increasing trust and adoption.