Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style

Authors: Julius von Kügelgen, Yash Sharma, Luigi Gresele, Wieland Brendel, Bernhard Schölkopf, Michel Besserve, Francesco Locatello

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform two main experiments. First, we numerically test our main result, Thm. 4.4, in a fullycontrolled, finite sample setting ( 5.1), using CL to estimate the entropy term in (5). Second, we seek to better understand the effect of data augmentations used in practice ( 5.2).
Researcher Affiliation Collaboration 1 Max Planck Institute for Intelligent Systems Tübingen 2 University of Cambridge 3 Tübingen AI Center, University of Tübingen 4 IMPRS for Intelligent Systems 5 Amazon
Pseudocode No No pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes Code available at: https://www.github.com/ysharma1126/ssl_identifiability
Open Datasets Yes We made the Causal3DIdent dataset publicly available at this URL.
Dataset Splits No The paper does not explicitly state specific training, validation, or test dataset splits (e.g., percentages or sample counts) within the provided text.
Hardware Specification No The paper mentions that compute and resources are 'Provided in Appendix D', but these details are not present in the provided text.
Software Dependencies No The paper mentions that software dependencies are 'Provided in Appendix D', but specific version numbers for key software components are not present in the provided text.
Experiment Setup Yes Experimental setup. We generate synthetic data as described in 3. We consider nc = ns = 5, with content and style latents distributed as c N(0, c) and s|c N(a + Bc, s), thus allowing for statistical dependence within the two blocks (via c and s) and causal dependence between content and style (via B). For f, we use a 3-layer MLP with Leaky Re LU activation functions. Experimental setup. For g, we train a convolutional encoder composed of a Res Net18 [46] and an additional fully-connected layer, with Leaky Re LU activation. As in Sim CLR [20], we use Info NCE with cosine similarity, and train on pairs of augmented examples ( x, x0). As nc is unknown and variable depending on the augmentation, we fix dim(ˆc) = 8 throughout.