A Probabilistic Model for Dirty Multi-task Feature Selection

Authors: Daniel Hernandez-Lobato, Jose Miguel Hernandez-Lobato, Zoubin Ghahramani

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Several experiments show that a model based on the new robust prior provides better predictive performance than other benchmark methods.
Researcher Affiliation Academia Daniel Hern andez-Lobato DANIEL.HERNANDEZ@UAM.ES Universidad Aut onoma de Madrid, Computer Science Department, Madrid, 28049, SpainJos e Miguel Hern andez-Lobato JMH@SEAS.HARVARD.EDU Harvard University, School of Engineering and Applied Sciences, Cambridge, MA 02138, USAZoubin Ghahramani ZOUBIN@ENG.CAM.AC.UK University of Cambridge, Department of Engineering, Cambridge CB2 1PZ, UK
Pseudocode No No pseudocode or algorithm blocks are provided in the paper.
Open Source Code Yes The complete details about EP are found in the supplementary material, alongside with an R implementation of the proposed method.
Open Datasets Yes The experimental protocol follows the DREAM 4 in silico challenge 2009. We consider the dataset described in (Barretina et al., 2012). We consider the problem of denoising the 256 256 house image used in (Titsias & L azaro-Gredilla, 2011)
Dataset Splits Yes We use 90% of the instances for training and 10% for testing.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory amounts) used for running the experiments are mentioned in the paper.
Software Dependencies No All methods described are implemented in the R language.
Experiment Setup Yes DMFS, STL and MFS need not fix any hyper-parameters since they infer them from the data using hyper-priors. Unless stated differently, in all probabilistic models we assume different levels of noise for each task when training. In DM and RMFL we choose hyper-parameters using a grid search guided by an inner cross-validation method.