On Causally Disentangled Representations

Authors: Abbavaram Gowtham Reddy, Benin Godfrey L, Vineeth N Balasubramanian8089-8097

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform an empirical study on state of the art disentangled representation learners using our metrics and dataset to evaluate them from causal perspective.
Researcher Affiliation Academia Abbavaram Gowtham Reddy, Benin Godfrey L, Vineeth N Balasubramanian Indian Institute of Technology Hyderabad, India cs19resch11002@iith.ac.in, benin.godfrey@cse.iith.ac.in, vineethnb@iith.ac.in
Pseudocode Yes The algorithms detailing the implementation of UC and CG metrics are provided in the Appendix.
Open Source Code Yes CANDLE dataset, code, and the Appendix of this work are made publicly available at https://causal-disentanglement.github.io/IITHCANDLE/.
Open Datasets Yes We introduce an image dataset called CANDLE (Causal ANalysis in Disentang Led r Epresentations) with 6 data generating factors along with both observed and unobserved confounders... CANDLE dataset, code, and the Appendix of this work are made publicly available at https://causal-disentanglement.github.io/IITHCANDLE/. ... d Sprites (Matthey et al. 2017), MPI3D (Gondal et al. 2019).
Dataset Splits No The paper mentions “Semi-supervision is provided by using labels for 10% of data points” but does not specify explicit training/validation/test splits (e.g., percentages or counts) for the datasets used in experiments.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU model, CPU type, memory) used for running the experiments.
Software Dependencies No The paper mentions “Blender (Community 2018)” for dataset creation and “open-source disentanglement library (Locatello et al. 2019)” for training models, but it does not provide specific version numbers for software dependencies or libraries.
Experiment Setup Yes Semi-supervision is provided by using labels for 10% of data points. Additional details on the experimental setup and qualitative results are provided in the Appendix. Our loss function w.r.t. dataset D = {xi}L i=1 is given by: LS S FVAE BB = L(Factor VAE) +λ i=1 ||xi wi ˆxi wi||2 2 (6) where ... λ is a hyperparameter...