Invariant Models for Causal Transfer Learning

Authors: Mateo Rojas-Carulla, Bernhard Schölkopf, Richard Turner, Jonas Peters

JMLR 2018 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present results on synthetic data sets and a gene deletion data set. ... In Figure 4, we are in the DG setting (thus, no labeled examples from T are observed at training). ... We compare our estimator to different methods, which are summarized in Table 2. ... We run 100 repetitions and plot the histograms of = E(βdom) E(βCS(cau )).
Researcher Affiliation Academia Mateo Rojas-Carulla EMAIL Max Planck Institute for Intelligent Systems T ubingen, Germany Department of Engineering Univ. of Cambridge, United Kingdom Bernhard Sch olkopf EMAIL Max Planck Institute for Intelligent Systems T ubingen, Germany Richard Turner EMAIL Department of Engineering Univ. of Cambridge, United Kingdom Jonas Peters EMAIL Department of Mathematical Sciences Univ. of Copenhagen, Denmark
Pseudocode Yes It is summarized in Algorithm 1, code is provided in https://github.com/mrojascarulla/causal_transfer_learning. Algorithm 1: Subset search Algorithm 2: Greedy subset search
Open Source Code Yes It is summarized in Algorithm 1, code is provided in https://github.com/mrojascarulla/causal_transfer_learning. Appendix B. The code to reproduce the experiments in the paper can be found in https://github.com/mrojascarulla/causal_transfer_learning.
Open Datasets Yes We apply our method to gene perturbation data provided by Kemmeren et al. (2014). This data set consists of the m-RNA expression levels of p = 6170 genes X1, . . . , Xp of the Saccharomyces cerevisiae (yeast). Kemmeren, K., Sameith, L., van de Pasch, L., Benschop, J., Lenstra, T., Margaritis, T., ODuibhir, E., Apweiler, E., van Wageningen, S., Ko, C., et al. Large-scale genetic perturbations reveal regulatory networks and an abundance of gene-specific repressors. Cell, 157(3):740 752, 2014.
Dataset Splits Yes Moreover, we use 40% of the training examples to fit the linear models in Algorithms 1 and 2, and the remaining data as validation. ... 4000 examples per training task are available for the top left and right plots, while only 1000 examples per task are available on the bottom because of computational reasons. ... for each gene in (XL1, . . . , XL10) for which an intervention is available, we put aside the example corresponding to this intervention from the training data ... we test all methods on the single intervention point which was put aside.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU models, CPU types, or memory) used for running its experiments.
Software Dependencies No The paper mentions methods like Lasso, HSIC, Levene test, and EM algorithm but does not specify version numbers for any programming languages, libraries, or software tools used.
Experiment Setup Yes For all experiments, we choose δ = 0.05 as a rejection level for the statistical test in Algorithms 1 and 2. Moreover, we use 40% of the training examples to fit the linear models in Algorithms 1 and 2, and the remaining data as validation. ... In MTL, ... we select the accepted set ˆS which leads to the smallest 5 fold cross validation error. ... If one has reasons to believe that the signal is sparse, ... one may use a variable selection technique such as the Lasso (Tibshirani, 1996) as a first step.