DCI-ES: An Extended Disentanglement Framework with Connections to Identifiability

Authors: Cian Eastwood, Andrei Liviu Nicolicioiu, Julius Von Kügelgen, Armin Kekić, Frederik Träuble, Andrea Dittadi, Bernhard Schölkopf

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, in our experiments ( 6), we use our extended framework to compare different representations on the MPI3D-Real (Gondal et al., 2019) and Cars3D (Reed et al., 2015) datasets, illustrating the practical usefulness of our E score through its strong correlation with downstream performance.
Researcher Affiliation Academia Cian Eastwood 1,2, Andrei Liviu Nicolicioiu 1, Julius von Kügelgen 1,3, Armin Keki c1, Frederik Träuble1, Andrea Dittadi1,4, and Bernhard Schölkopf 1 1Max Planck Institute for Intelligent Systems, Tübingen, Germany 2School of Informatics, University of Edinburgh 3Department of Engineering, University of Cambridge 4Technical University of Denmark
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Further details can be found in our open-source code: https://github.com/andreinicolicioiu/DCI-ES.
Open Datasets Yes Data. We perform our analysis of loss-capacity curves on the MPI3D-Real (Gondal et al., 2019) and Cars3D (Reed et al., 2015) datasets.
Dataset Splits Yes We split the data into training, validation and test sets of size 295k, 16k, and 726k respectively for MPI3D-Real and 12.6k, 1.4k, 3.4k for Cars3d. We use the validation split for hyperparameter selection and report results on the test split.
Hardware Specification No The paper describes the probes and training process but does not specify any hardware details like GPU or CPU models used for the experiments.
Software Dependencies No The paper mentions software components like 'Adam optimizer' and 'SAGE' but does not provide specific version numbers for any software or libraries.
Experiment Setup Yes We split the data into training, validation and test sets of size 295k, 16k, and 726k respectively for MPI3D-Real and 12.6k, 1.4k, 3.4k for Cars3d. We use the validation split for hyperparameter selection and report results on the test split. We train MLP probes using the Adam (Kingma & Ba, 2015) optimizer for 100 epochs. We use mean-square error and cross-entropy losses for continuous and discrete factors zj, respectively. To compute Ej, we use the baseline losses of E[zj] and a random classifier for continuous and discrete zj, respectively.