Replacing the conditional GP prior inside sparse variational inference with a Gaussian sharing its mean but using diagonal, analytically optimal variance corrections yields a strictly tighter evidence lower bound at unchanged O(NM^2) cost.
σ2 σ2 f ℓ2 Exact GP 0.0715 0.712 0.597 SVGP-new 0.087 0.485 0.615 SVGP 0.108 0.331 0.617 D
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
ACCEPT 1representative citing papers
citing papers explorer
-
New Bounds for Sparse Variational Gaussian Processes
Replacing the conditional GP prior inside sparse variational inference with a Gaussian sharing its mean but using diagonal, analytically optimal variance corrections yields a strictly tighter evidence lower bound at unchanged O(NM^2) cost.