Embrace the Gap: VAEs Perform Independent Mechanism Analysis

Authors: Patrik Reizinger, Luigi Gresele, Jack Brady, Julius von Kügelgen, Dominik Zietlow, Bernhard Schölkopf, Georg Martius, Wieland Brendel, Michel Besserve

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

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments on synthetic and image data, we show that VAEs uncover the true latent factors when the data generating process satisfies the IMA assumption.
Researcher Affiliation Collaboration 1University of Tübingen, Germany 2Max Planck Institute for Intelligent Systems, Tübingen, Germany 3University of Cambridge, Cambridge, United Kingdom 4Amazon Web Services, Tübingen, Germany
Pseudocode No The paper describes its methods in detail using mathematical notation and prose, but it does not include any explicitly labeled pseudocode blocks or algorithms.
Open Source Code Yes Code available at: github.com/rpatrik96/ima-vae
Open Datasets Yes Experimental setup (image). We train a VAE (not β-VAE) with a factorized Gaussian posterior and Beta prior on a Sprites image dataset generated using the spriteworld renderer [66] with a Beta ground truth distribution.
Dataset Splits No The paper describes the datasets and training process, but it does not explicitly state the specific percentages or counts used for training, validation, and test splits. It mentions using 60,000 source samples and training for a certain number of seeds and epochs, but without explicit split ratios for validation.
Hardware Specification Yes All experiments were run on NVIDIA A100 GPUs.
Software Dependencies Yes Our implementation is in PyTorch 1.10.1.
Experiment Setup Yes We use a learning rate of 1e-3, a batch size of 256 for 3D experiments and 128 for 2D, and train for 500 epochs with the Adam optimizer (β1 = 0.9, β2 = 0.999). The 3-layer MLP encoder and decoder have dimensions of [3, 256, 256, 3] and use smooth Leaky ReLU nonlinearities.