CITRIS: Causal Identifiability from Temporal Intervened Sequences
Authors: Phillip Lippe, Sara Magliacane, Sindy Löwe, Yuki M Asano, Taco Cohen, Stratis Gavves
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments on 3D rendered image sequences, CITRIS outperforms previous methods on recovering the underlying causal variables. Moreover, using pretrained autoencoders, CITRIS can even generalize to unseen instantiations of causal factors, opening future research areas in sim-to-real generalization for causal representation learning. We evaluate CITRIS on two video datasets, and compare it to common disentanglement methods. |
| Researcher Affiliation | Collaboration | 1QUVA Lab, University of Amsterdam, Amsterdam, The Netherlands 2Institute of Informatics, University of Amsterdam, Amsterdam, The Netherlands 3MIT-IBM Watson AI Lab 4UvA-Bosch Delta Lab, University of Amsterdam, Amsterdam, The Netherlands 5Qualcomm AI Research, Amsterdam, The Netherlands. Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc. |
| Pseudocode | No | The paper describes algorithms and models using mathematical formulations and architectural diagrams, but does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | We include further details for reproducibility in Appendix C, and make our code publicly available.1 1https://github.com/phlippe/CITRIS To ensure reproducibility, we publish the code for all models used in this paper at https://github.com/phlippe/ CITRIS. |
| Open Datasets | Yes | The complete datasets of this paper are released with corresponding licenses, and links to those datasets are available in the code repository. The creation of the Temporal Causal3DIdent dataset closely followed the setup of von Kugelgen et al. (2021); Zimmermann et al. (2021). |
| Dataset Splits | No | To train the MLP, we split the test dataset into two subsets: one for training the MLP (40% of the dataset), and one for measuring the correlation metrics on (60% of the dataset). |
| Hardware Specification | Yes | Finally, all experiments in this paper were performed on a single NVIDIA Titan RTX GPU with a 6-core CPU. |
| Software Dependencies | No | All models are implemented in the Deep Learning framework Py Torch (Paszke et al., 2019) and Py Torch Lightning (Falcon & The Py Torch Lightning team, 2019). |
| Experiment Setup | Yes | We provide an overview of the used hyperparameters in Table 6 (VAE models), Table 7 (Autoencoder training), and Table 8. In general, we use the Adam optimizer (Kingma & Ba, 2015) with a learning rate of 1e-3. We use a Cosine Warmup learning rate scheduler with 100 steps warmup. |