Sparse Variational Student-t Processes
Published in Proceedings of the AAAI Conference on Artificial Intelligence, 38(14), 2024
SVTP extends the sparse inducing-point framework from Gaussian Processes to Student-t Processes (TPs). We derive a tractable evidence lower bound using the conditional Student-t distribution at inducing points and propose two inference algorithms — SVTP-UB (Jensen’s inequality upper bound) and SVTP-MC (Monte Carlo) — together with a theoretical analysis of when SVTP out-performs SVGP. Empirically, SVTP delivers substantial gains on outlier-contaminated and heavy-tailed datasets.
Recommended citation: Xu, J. and Zeng, D. (2024). Sparse variational student-t processes. Proceedings of the AAAI Conference on Artificial Intelligence, 38(14), pp.16156-16163.
Download Paper