Latent Gaussian Processes for Distribution Estimation of Multivariate Categorical Data
Authors: Yarin Gal, Yutian Chen, Zoubin Ghahramani
ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show empirically that the model outperforms its linear and discrete counterparts in imputation tasks of sparse data. We experimentally show the advantages of using non-linear transformations for the latent space. We evaluate the categorical latent Gaussian process (CLGP), comparing it to existing models for multivariate categorical distribution estimation. |
| Researcher Affiliation | Academia | Yarin Gal YG279@CAM.AC.UK Yutian Chen YC373@CAM.AC.UK Zoubin Ghahramani ZOUBIN@CAM.AC.UK University of Cambridge |
| Pseudocode | No | The paper states 'Python code for the model and inference is given in the appendix'. While this implies implementation details, the main text of the paper does not contain any structured pseudocode or clearly labeled algorithm blocks. The appendix is not provided as part of the document for analysis. |
| Open Source Code | Yes | Python code for the model and inference is given in the appendix, and available at http://github.com/yaringal/CLGP |
| Open Datasets | Yes | We use the Wisconsin breast cancer dataset for which obtaining samples is a long and expensive process3. Obtained from the UCI repository. obtained from the START Terrorism Data Archive Dataverse4. Obtained from the Harvard Dataverse Network thedata.harvard.edu/dvn/dv/start/study/StudyPage.xhtml?studyId=71190. The dataset, Binary Alphadigits... Obtained from http://www.cs.nyu.edu/roweis/data.html |
| Dataset Splits | Yes | We use three quarters of the dataset for training and leave the rest for testing, averaging the testset perplexity on three repetitions of the experiment. We used 1000 cases for training and 282 for test. Each class is then randomly split to 30 training and 9 test examples. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used for running the experiments (e.g., GPU models, CPU types, memory specifications, or cloud instance types). It only mentions 'Theano' as a symbolic differentiation library, which can run on various hardware, but no specific hardware setup is listed. |
| Software Dependencies | No | The paper mentions 'Theano' (Bergstra et al., 2010) for symbolic differentiation and implies the use of Python ('Python code for the model'). However, it does not provide specific version numbers for Theano, Python, or any other libraries or frameworks used, which are necessary for reproducible software dependencies. |
| Experiment Setup | Yes | For the following experiments, both the linear and non-linear models were initialised with a 2D latent space. The mean values of the latent points, mn, were initialised at random following a standard normal distribution. The mean values of the inducing outputs (µdk) were initialised with a normal distribution with standard deviation 10 2. We initialise the standard deviation of each latent point (sn) to 0.1, and initialise the length-scales of the ARD RBF covariance function to 0.1. We then optimise the variational distribution for 500 iterations. We repeat the same experiment set-up as before, with a 6 dimensional latent space, 100 inducing points, 5 samples to estimate the lower bound, and running the optimiser for 1000 iterations. |