Bayesian Learning via Q-Exponential Process

Authors: Shuyi Li, Michael O'Connor, Shiwei Lan

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare GP, Besov and Q-EP in modeling functional data, reconstructing images and solving inverse problems and demonstrate the advantage of our proposed methodology. In this section, we compare GP, Besov and Q-EP by modeling time series (temporal), reconstructing images (spatial) from computed tomography and solving a (spatiotemporal) inverse problem (Appendix C.4). These numerical experiments demonstrate that our proposed Q-EP enables faster convergence in obtaining a better maximum a posterior (MAP) estimate.
Researcher Affiliation Academia Shuyi Li Michael O Connor Shiwei Lan School of Mathematical & Statistical Sciences Arizona State University, Tempe, AZ 85287 slan@asu.edu
Pseudocode Yes Algorithm 1 White-noise Preconditioed Crank-Nicolson (wn-p CN) for Q-EP Prior Models
Open Source Code Yes All the computer codes are publicly available at https://github.com/lanzithinking/Q-EXP.
Open Datasets Yes In addition, we also consider two real data sets of Tesla and Google stock prices in 2022. We first consider the Shepp Logan phantom, a standard test image created by Shepp and Logan in [42] to model a human head and to test image reconstruction algorithms. Finally, we apply these methods to CT scans of a human cadaver and torso from the Visible Human Project [1].
Dataset Splits No The paper specifies training and testing splits for its datasets, but does not explicitly mention a distinct validation set or its proportion.
Hardware Specification No The paper does not provide specific details on the hardware used for running experiments, such as GPU/CPU models or cloud instance types.
Software Dependencies No The paper mentions algorithms used (e.g., BFGS, MCMC) but does not provide specific version numbers for software dependencies or libraries.
Experiment Setup Yes For both GP and Q-EP, we adopt the Matérn kernel with ν = 1 2, σ2 = 1, ρ = 0.5 and s = 1: C(t,t ) = σ2 21 ν Γ(ν) wνKν(w), w = 2ν( t t /ρ)s. In both Besov and Q-EP, we set q = 1. The generated sinogram is then added by noise with signal noise ratio SNR = Au / ε = 100. We set γ = 1 and δ = 8 in this example. For Besov, we adopt 2d Fourier basis... and truncate the series (1) for the first L = 1000 terms.