Decoupled Variational Gaussian Inference
Authors: Mohammad Emtiyaz Khan
NeurIPS 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the advantages of our approach on a binary GP classification problem. We model the binary data using Bernoulli-logit likelihoods. We apply this model to a subproblem of the USPS digit data [18]. |
| Researcher Affiliation | Academia | Mohammad Emtiyaz Khan Ecole Polytechnique F ed erale de Lausanne (EPFL), Switzerland emtiyaz@gmail.com |
| Pseudocode | Yes | Algorithm 1 Linearly constrained Lagrangian (LCL) method for VG approximation |
| Open Source Code | No | The paper states: 'In future, we plan to have an efficient implementation of this method and demonstrate that this enables variational inference to scale to large data.', indicating future release, not current availability. No explicit link or statement of immediate code release was found. |
| Open Datasets | Yes | We apply this model to a subproblem of the USPS digit data [18]. |
| Dataset Splits | No | The paper mentions using a subproblem of the USPS digit data with 'a total of 1540 data examples' and subsampling randomly, but does not specify exact training, validation, or test splits (e.g., percentages, counts, or references to predefined splits). |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions 'L-BFGS method for optimization (implemented in min Func by Mark Schmidt)', but does not provide specific version numbers for 'minFunc' or any other software dependencies. |
| Experiment Setup | Yes | We set µ = 0 and use a squared-exponential kernel, for which the (i, j)th entry of Σ is defined as: Σij = σ2 exp[ 1 2||xi xj||2/s] where xi is i th feature. We show results for log(σ) = 4 and log(s) = 1... All algorithms were stopped when the subsequent changes in the lower bound value of Eq. 5 were less than 10 4. |