Robustly Disentangled Causal Mechanisms: Validating Deep Representations for Interventional Robustness

Authors: Raphael Suter, Djordje Miladinovic, Bernhard Schölkopf, Stefan Bauer

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Section 6 provides experimental evidence in a standard disentanglement benchmark dataset supporting the need of a robustness based disentanglement criterion. Our extensive experiments on a standard benchmark dataset show that our robustness based validation is able to discover vulnerabilities of deep representations that have been undetected by existing work.
Researcher Affiliation Academia 1Department of Computer Science, ETH Zurich, Switzerland 2MPI for Intelligent Systems, T ubingen, Germany.
Pseudocode Yes Algorithm 1 EMPIDA Estimation 1: Input: 2: dataset D = {(x(i), g(i))}i=1,...,N 3: trained encoder E 4: subsets of factors L {1, . . . , K0} and I, J {1, . . . , K} 5: Preprocessing: 6: encode all samples to obtain {z(i) = E(x(i)) : i = 1, . . . , N} 7: estimate p(g(i)) and p(g(i)\ (I[J)) 8i from relative frequencies in D 8: Estimation: 9: find all realizations of GI in D: {g(k)I , k = 1, . . . , NI} 10: partition the dataset according to those realizations: I := {(x, g) 2 D s.t. g I = g(k)I } 11: for k = 1, . . . , NI do 12: estimate mean E[ZL|do(GI g(k)I )] using Eq. (7) and samples D(k)I 13: partition D(k)I according to realizations of GJ: D(k,l)I,J := {(x, g) 2 D(k)I s.t. g J = g(l)J } 14: initialize mpida(k) 0.0 15: for l = 1, . . . , N (k) 16: meanint E[ZL|do(GI g(k)I , GJ g(l)J )] using Eq. (7) and samples D(k,l)I,J for estimation 17: compute pida d(mean, meanint) 18: update mpida(k) max (mpida(k), pida) 19: end for 20: end for 21: Return empida PNI I | |D| mpida(k)
Open Source Code Yes For future extensions and applications our work is added to the disentanglement lib of Locatello et al. (2018).
Open Datasets Yes In many benchmark datasets for disentanglement (e.g. dsprites) the observations are obtained noise-free and the dataset contains all possible combinations of generative factors exactly once.
Dataset Splits No The paper mentions using 'observational data' and 'benchmark datasets' like 'dsprites' but does not specify the exact train/validation/test splits used for their experiments within the provided text.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory, or cloud instances) used for running experiments are provided in the paper.
Software Dependencies No The paper mentions various VAE implementations (e.g., classic VAE, β-VAE, DIP-VAE) and refers to the 'disentanglement lib of Locatello et al. (2018)' but does not provide specific version numbers for any software dependencies.
Experiment Setup No The paper states 'we used each method with the parameter settings that were indicated in the original publications (details are given in Appendix D)', but Appendix D is not included in the provided text, so specific experimental setup details are not accessible.