Sparse Variational Student-t Processes for Heavy-Tailed Modeling
Published in IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2026
A principled sparse framework for Student-t Processes (TPs) extending SVTP (AAAI 2024). We derive the Fisher information matrix of the multivariate Student-t variational distribution and show its components have a compact expression in terms of the beta function (the “beta link”), giving the first tractable method for natural-gradient learning of sparse TPs. The result is substantially faster convergence and lower prediction error on UCI and Kaggle datasets, and the framework scales to datasets with over 200,000 samples while preserving TP’s outlier robustness.
Recommended citation: Xu, J., Zeng, D. and Paisley, J. (2026). Sparse Variational Student-t Processes for Heavy-Tailed Modeling. IEEE Transactions on Neural Networks and Learning Systems, pp.1-14, doi:10.1109/TNNLS.2026.3673350.
Download Paper