Variational Implicit Processes

Authors: Chao Ma, Yingzhen Li, Jose Miguel Hernandez-Lobato

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that VIPs return better uncertainty estimates and lower errors over existing inference methods for challenging models such as Bayesian neural networks, and Gaussian processes.
Researcher Affiliation Collaboration 1Department of Engineering, University of Cambridge, Cambridge, UK 2Microsoft Research Cambridge, Cambridge, UK.
Pseudocode Yes Algorithm 1 Variational Implicit Processes (VIP)
Open Source Code No The paper does not contain an explicit statement providing concrete access to source code for the methodology described.
Open Datasets Yes We use an IP with a Bayesian neural network (1-10-10-1 architecture) as the prior. We use α = 0 for the wake-step training. We also compare VIP with the exact full GP with optimized compositional kernel (RBF+Periodic), and another BNN with identical architecture but trained using variational dropout (VDO) with dropout rate p = 0.99 and length scale l = 0.001. The (hyper-)parameters are optimized using 500 epochs (batch training) with Adam optimizer (learning rate = 0.01). ... We compare the VIP (α = 0) with a variationally sparse GP (SVGP, 100 inducing points), an exact GP and VDO on the solar irradiance dataset (Lean et al., 1995). ... using real-world multivariate regression datasets from the UCI data repository (Lichman et al., 2013). ... Harvard Clean Energy Project Data, the world s largest materials high-throughput virtual screening effort (Hachmann et al., 2014).
Dataset Splits Yes The observational noise variance for VIP and VDO is tuned over a validation set, as detailed in Appendix F.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions 'Adam optimizer' but does not specify software names with version numbers for its dependencies.
Experiment Setup Yes The (hyper-)parameters are optimized using 500 epochs (batch training) with Adam optimizer (learning rate = 0.01).