Problem
Research questions and friction points this paper is trying to address.
Accelerating inference for non-conjugate Gaussian processes
Reducing computation cost while increasing uncertainty
Improving scalability for large datasets in NCGPs
Accelerating Non-Conjugate Gaussian Processes By Trading Off Computation For Uncertainty
🤖 AI Summary
Exact inference for non-conjugate Gaussian processes (NCGPs) is computationally prohibitive on large-scale data, while existing approximations suffer from unquantified errors that impair uncertainty calibration. To address this, we propose an error-aware parallel iterative variational inference framework. Our key contribution is the first explicit modeling of approximation error as a stochastic variable, enabling controlled trade-offs between computational cost and predictive uncertainty. By integrating parallelized low-rank approximations with adaptive information compression, the framework supports computation reuse and overcomes scalability bottlenecks inherent in traditional NCGP inference. Experiments on large-scale classification tasks demonstrate speedups of several- to ten-fold in posterior inference, substantial reductions in memory and time overhead, and—critically—preserved uncertainty calibration performance.
📝 Abstract
Non-conjugate Gaussian processes (NCGPs) define a flexible probabilistic framework to model categorical, ordinal and continuous data, and are widely used in practice. However, exact inference in NCGPs is prohibitively expensive for large datasets, thus requiring approximations in practice. The approximation error adversely impacts the reliability of the model and is not accounted for in the uncertainty of the prediction. We introduce a family of iterative methods that explicitly model this error. They are uniquely suited to parallel modern computing hardware, efficiently recycle computations, and compress information to reduce both the time and memory requirements for NCGPs. As we demonstrate on large-scale classification problems, our method significantly accelerates posterior inference compared to competitive baselines by trading off reduced computation for increased uncertainty.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Iterative methods modeling approximation error explicitly
Parallel computing hardware efficient computation recycling
Information compression reducing time and memory
🔎 Similar Papers
No similar papers found.
