Visual Representation Learning Does Not Generalize Strongly Within the Same Domain

Authors: Lukas Schott, Julius Von Kügelgen, Frederik Träuble, Peter Vincent Gehler, Chris Russell, Matthias Bethge, Bernhard Schölkopf, Francesco Locatello, Wieland Brendel

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

Reproducibility Variable Result LLM Response
Research Type Experimental In total, we train and test 2000+ models and observe that all of them struggle to learn the underlying mechanism regardless of supervision signal and architectural bias.
Researcher Affiliation Collaboration 1University of Tübingen, 2Max Planck Institute for Intelligent Systems, Tübingen 3University of Cambridge, 4Amazon Web Services
Pseudocode No The paper describes the experimental setup and training details in textual paragraphs (e.g., Section 4, 5, and H.3), but does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes To this end, all data sets and evaluation scripts are released alongside a leaderboard on Git Hub. 1https://github.com/bethgelab/In Domain Generalization Benchmark
Open Datasets Yes We consider datasets with images generated from a set of discrete Factors of Variation (Fo Vs)... d Sprites (Matthey et al., 2017), ... Shapes3D (Kim & Mnih, 2018), ... MPI3D (Gondal et al., 2019)...
Dataset Splits No We further control all considered splits and datasets such that 30% of the available data is in the training set Dtr and the remaining 70% belong to the test set Dte. The paper explicitly defines train and test splits but does not specify a separate, dedicated validation set percentage or size for reproduction.
Hardware Specification Yes All models are run on the NVIDIA T4 Tensor Core GPUs on the AWS g4dn.4xlarge instances with an approximate total compute of 20 000 GPUh.
Software Dependencies Yes All models are implemented using Py Torch 1.7.
Experiment Setup Yes All fully supervised models are trained with the same training scheme. We use the Adam optimizer with a learning rate of 0.0005. ... We train each model with three random seeds for 500, 000 iterations with a batch size of b = 64. As a loss function, we consider the mean squared error MSE...