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 and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Dependent relevance determination for smooth and structured sparse regression
Authors: Anqi Wu, Oluwasanmi Koyejo, Jonathan Pillow
JMLR 2019 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We finally show substantial improvements over comparable methods for both simulated and real datasets from brain imaging. Keywords: Bayesian nonparametric, Sparsity, Structure learning, Gaussian Process, f MRI |
| Researcher Affiliation | Academia | Anqi Wu EMAIL Princeton Neuroscience Institute Princeton University Princeton, NJ 08544, USA Oluwasanmi Koyejo EMAIL Beckman Institute for Advanced Science and Technology Department of Computer Science University of Illinois at Urbana-Champaign Urbana, Illinois, 61801, USA Jonathan Pillow EMAIL Princeton Neuroscience Institute Princeton University Princeton, NJ 08544, USA |
| Pseudocode | Yes | Algorithm 1 A two-stage convex relaxation method for DRD Laplace approximation Algorithm 2 Evidence optimization using decoupled Laplace approximation |
| Open Source Code | Yes | The code and simulated results are available online4. 4. https://github.com/waq1129/DRD.git |
| Open Datasets | Yes | We finally show substantial improvements over comparable methods for both simulated and real datasets from brain imaging. We first considered the regression problem of decoding gains and losses from f MRI measurements recorded in a gambling task (Tom et al., 2007, 2011). Next we considered the problem of predicting a subject s age from a measured map of gray-matter concentration, using data from the Open Access Series of Imaging Studies (OASIS) (Marcus et al., 2007). We used a popular f MRI dataset from a study on face and object representation in human ventral temporal cortex (Haxby et al., 2001). |
| Dataset Splits | Yes | We varied training set size from n = 100 to 400 and kept a fixed test size of 100 samples. (We noted that even with n = 400 samples, the problem resides in the n < p small-sample regime). We repeated each experiment 5 times. To assess the performance, we varied the train-test ratio in number of subjects from 9:7 to 13:3. We performed 10 different random train-test splits for each ratio. To assess the performance, we varied the training ratio from 0.4 to 0.8 out of the 403 subjects, and averaged over 5 random splits for each ratio. We divided 12 sessions of data per subject into train-test splits of 5:7, 6:6 and 7:5. |
| Hardware Specification | No | The paper does not provide specific hardware details used for running its experiments. |
| Software Dependencies | No | lasso (Tibshirani, 1996), using Least Angle Regression (LARS) implemented by glmnet1; We computed total variation l1 (TV-L1) and graph net (Graph Net) estimates using the Nilearn2 package (Abraham et al., 2014). SCAD was implemented by Sparse Reg3 (Zhou and Gaines, 2017). The paper mentions software packages like glmnet, Nilearn, and Sparse Reg, but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | We sampled a p = 4000 dimensional weight vector w from the smooth-DRD prior (see Fig. 1), with hyperparameters GP mean b = 8, GP length scale l = 100, GP marginal variance ρ = 36, smoothness length scale δ = 50, measurement noise variance σ2 = 5. The hyper-hyperparameters in the MCMC methods (eq. 47) were set to: mn = 2, σ2 n = 5, mb = 10, σ2 b = 8, aρ = 4, bρ = 5, al = 4, bl = 25, aδ = 4, bδ = 25. To ensure the accuracy of the Laplace approximation, in each iteration t, we optimize eq. (39) over a restricted region of the hyperparameter space around the previous hyperparameter setting θt 1, which allows varying within 20% of its current value on each iteration in our experiments. For the Laplace method, we used a stopping criterion that the change in w was less than 0.0001. For the MCMC method, we assessed burn-in using a criterion on the relative change in w, and then collected 100 posterior samples. |