Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Bayesian Inference for Spatio-temporal Spike-and-Slab Priors
Authors: Michael Riis Andersen, Aki Vehtari, Ole Winther, Lars Kai Hansen
JMLR 2017 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we demonstrate the proposed model using numerical experiments based on both synthetic and real data sets. |
| Researcher Affiliation | Academia | Michael Riis Andersen EMAIL Aki Vehtari EMAIL Helsinki Institute for Information Technology HIIT Department of Computer Science, Aalto University P.O. Box 15400, FI-00076, Finland Ole Winther EMAIL Lars Kai Hansen EMAIL Department of Applied Mathematics and Computer Science Technical University of Denmark DK-2800 Kgs. Lyngby, Denmark |
| Pseudocode | Yes | Algorithm 1: Proposed algorithm for approximating the joint posterior distribution over X, Z and Γ conditioned on Y using parallel EP. |
| Open Source Code | Yes | The code is available at https://github.com/MichaelRiis/SSAS. |
| Open Datasets | Yes | In this experiment, we apply the structured spike-and-slab model with Gaussian likelihood to an application of compressed sensing (Donoho, 2006) of numerical characters from the MNIST data set (Le Cun et al., 1998; Hern andez-Lobato et al., 2013). In the final experiment, we apply the proposed method to an EEG source localization problem (Baillet et al., 2001), where the objective is to infer the locations of the active sources on the cortical surface of the human brain based on electroencephalogram (EEG) measurements. The data set is publicly available and the experimental paradigm is described in (Henson et al., 2003). |
| Dataset Splits | Yes | The data set consists of 695 and 1022 utterances of the vowels aa and ao, respectively, along with their corresponding labels. The response variable in this experiment is binary and therefore we use the probit model rather than the Gaussian observation model. ... We use N = 150 examples for training and the remaining 1567 examples for testing. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory details) used for running its experiments. |
| Software Dependencies | Yes | We used the SPM8 software (Ashburner et al., 2010). We used the implementation in scikit-learn toolbox (Pedregosa et al., 2011). We used the implementation in SPAMS toolbox (Jenatton et al., 2010; Mairal et al., 2010). |
| Experiment Setup | Yes | The noise variance is fixed to the true value and the remaining hyperparameters are fixed ν = 0, τ = 1, κ2 1 = 5. We fixed the expected sparsity to K = 1/4D = 125 by choosing the prior mean of γ to ν = φ^-1(1/4(1 + κ^2_1)). The regularization parameters for the ridge regression is fixed to λ = 10^-3 for all runs. We use a squared exponential kernel with a single lengthscale defined on the 2D image grid to encourage the neighbourhood structure expected in the images. We impose a Gaussian prior distribution on ν0 with zero mean and variance κ2 1, that is ν0 N(0, κ2 1) and integrate over ν0 analytically to get the kernel function k(i, j) = κ2 1 + κ2 2 exp(di dj 2 / (2κ2 3)). ... we fix the prior mean and variance of the slab component to a standardized Gaussian with ρ0 = 0 and τ0 = 1. Thus, the hyperparameters to be learned are Ω= {κ1, κ2, κ3}. |