Probabilistic size-and-shape functional mixed models

Authors: Fangyi Wang, Karthik Bharath, Oksana Chkrebtii, Sebastian Kurtek

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our numerical experiments demonstrate utility of the proposed model, and superiority over the current state-of-the-art. We present posterior inference results from the model described in Section 3 for simulated and real data.
Researcher Affiliation Academia Fangyi Wang Department of Statistics The Ohio State University Columbus, OH, 43210 wang.15022@osu.edu Karthik Bharath School of Mathematical Sciences University of Nottingham Nottingham, UK, NG7 2RD Karthik.Bharath@nottingham.ac.uk Oksana Chkrebtii Department of Statistics The Ohio State University Columbus, OH, 43210 oksana@stat.osu.edu Sebastian Kurtek Department of Statistics The Ohio State University Columbus, OH, 43210 kurtek.1@stat.osu.edu
Pseudocode Yes The detailed MCMC algorithm is given in Algorithm 1.
Open Source Code Yes The supplementary material include data and code for reproducibility purposes.
Open Datasets Yes Consider a sample of functions from the much-studied Berkeley growth data in Figure 1(a)... We now consider application of the proposed modeling framework to (i) Berkeley growth rate functions (n = 93) [Srivastava et al., 2011b], and (ii) PQRST complexes (n = 40) [Kurtek et al., 2013].
Dataset Splits No The paper conducts numerical experiments on simulated and real data but does not specify any training, validation, or test splits. It mentions MCMC burn-in periods but this relates to sampling, not data partitioning.
Hardware Specification Yes To evolve the MCMC algorithm in MATLAB R2021a for 300, 000 iterations yielding 100, 000 posterior samples after burn-in, on a computing server with 6 parallel Intel(R) Xeon(R) CPUs with 20GB of memory, the computing time is approximately 93 and 111 minutes, respectively, under PMs 1 and 2 on Γ, based on n = 30 functions discretized at T = 50 points.
Software Dependencies Yes To evolve the MCMC algorithm in MATLAB R2021a for 300, 000 iterations yielding 100, 000 posterior samples after burn-in...
Experiment Setup Yes We use Bf = 6 modified Fourier basis functions for µ and Br = 6 B-spline basis functions for each vi. The ground truth variances are σ2 c = 0.25 and σ2 = 0.1. ... For a, σ2 c and σ2, we use weakly informative prior distributions: a MVN(0, 10000IBf ), σ2 c IG(0.01, 0.01), σ2 IG(0.01, 0.01)... We use θγ = 30 in all of our numerical experiments. ... For all examples in this section, we use N = 100, 000 with a burn-in period of 200, 000 iterations.