{"id":15783654,"url":"https://github.com/deepmancer/rss-training-iclr2024","last_synced_at":"2026-01-11T02:38:46.589Z","repository":{"id":231843958,"uuid":"782844669","full_name":"deepmancer/rss-training-iclr2024","owner":"deepmancer","description":"This is the official repository for the ICLR 2024 paper Out-Of-Domain Unlabeled Data Improves Generalization.","archived":false,"fork":false,"pushed_at":"2024-09-02T14:32:19.000Z","size":32460,"stargazers_count":4,"open_issues_count":0,"forks_count":0,"subscribers_count":1,"default_branch":"main","last_synced_at":"2024-10-11T20:02:04.619Z","etag":null,"topics":["adversarial-learning","conference-paper","distribution-shift","distributionally-robust-optimization","generalization-error","iclr","iclr2024","out-of-distribution","out-of-distribution-generalization","out-of-domain","robust-optimization","robustness","self-supervised-learning","semi-supervised-learning","unsupervised-learning"],"latest_commit_sha":null,"homepage":"https://openreview.net/forum?id=Bo6GpQ3B9a","language":"TeX","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"apache-2.0","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/deepmancer.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null}},"created_at":"2024-04-06T07:40:09.000Z","updated_at":"2024-09-14T10:51:23.000Z","dependencies_parsed_at":"2024-10-04T20:00:38.391Z","dependency_job_id":"fafad840-1c72-4925-8bce-5692e806308d","html_url":"https://github.com/deepmancer/rss-training-iclr2024","commit_stats":{"total_commits":43,"total_committers":3,"mean_commits":"14.333333333333334","dds":"0.13953488372093026","last_synced_commit":"a8f9a70a31b3119d5b071c2d34218b0fc1e46b6b"},"previous_names":["alirezaheidari-cs/rss-training","deepmancer/rss-training","deepmancer/rss-training-iclr2024"],"tags_count":0,"template":false,"template_full_name":null,"repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/deepmancer%2Frss-training-iclr2024","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/deepmancer%2Frss-training-iclr2024/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/deepmancer%2Frss-training-iclr2024/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/deepmancer%2Frss-training-iclr2024/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/deepmancer","download_url":"https://codeload.github.com/deepmancer/rss-training-iclr2024/tar.gz/refs/heads/main","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":246500788,"owners_count":20787754,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["adversarial-learning","conference-paper","distribution-shift","distributionally-robust-optimization","generalization-error","iclr","iclr2024","out-of-distribution","out-of-distribution-generalization","out-of-domain","robust-optimization","robustness","self-supervised-learning","semi-supervised-learning","unsupervised-learning"],"created_at":"2024-10-04T20:00:27.810Z","updated_at":"2026-01-11T02:38:46.576Z","avatar_url":"https://github.com/deepmancer.png","language":"TeX","funding_links":[],"categories":[],"sub_categories":[],"readme":"# Out-Of-Domain Unlabeled Data Improves Generalization - ICLR 2024 (Spotlight)\n\n[![arXiv Paper](https://img.shields.io/badge/arXiv-Paper-\u003cCOLOR\u003e.svg)](https://arxiv.org/abs/2310.00027)\n[![Poster PDF](https://img.shields.io/badge/Poster-PDF-87CEEB)](https://iclr.cc/media/PosterPDFs/ICLR%202024/19202.png?t=1712876187.1666338)\n[![Presentation](https://img.shields.io/badge/Presentation-ICLR%202024-FFA500)](https://iclr.cc/virtual/2024/poster/19202)\n[![OpenReview Discussion](https://img.shields.io/badge/OpenReview-Discussion-B762C1)](https://openreview.net/forum?id=Bo6GpQ3B9a)\n\n\u003cdetails\u003e\n    \u003csummary\u003e📜 Click for Abstract\u003c/summary\u003e\n\nWe propose a novel framework for incorporating unlabeled data into semi-supervised classification problems, where scenarios involving the minimization of either:\n\n- *i)* adversarially robust, or \n- *ii)* non-robust loss functions \n\nhave been considered. Notably, we allow the unlabeled samples to deviate slightly (in the total variation sense) from the in-domain distribution. The core idea behind our framework is to combine **Distributionally Robust Optimization (DRO)** with **Self-Supervised Learning (SSL)**. As a result, we also leverage **efficient polynomial-time algorithms** for the training stage.\n\nFrom a theoretical standpoint, we apply our framework to the classification problem of a mixture of two Gaussians in $\\mathbb{R}^d$, where, in addition to the $m$ independent and labeled samples from the true distribution, a set of $n$ (usually with $n \\gg m$) out-of-domain and unlabeled samples are also provided.\n\nUsing only the labeled data, it is known that the generalization error can be bounded by:\n\n$$\\propto \\left(\\frac{d}{m}\\right)^{1/2}.$$\n\nHowever, using our method on both isotropic and non-isotropic Gaussian mixture models, one can derive a new set of analytically explicit and non-asymptotic bounds which show substantial improvement in the generalization error compared to ERM.\n\nOur results underscore two significant insights:\n\n1. Out-of-domain samples, even when unlabeled, can be harnessed to narrow the generalization gap, provided that the true data distribution adheres to a form of the *\"cluster assumption\"*.\n2. The semi-supervised learning paradigm can be regarded as a special case of our framework when there are no distributional shifts.\n\nWe validate our claims through experiments conducted on a variety of synthetic and real-world datasets.\n\n\u003c/details\u003e\n\n---\n\n## 💡 Methodology: Robust Self-Supervised (RSS) Training\n\nThe **Robust Self-Supervised (RSS) Training framework** enhances the ERM loss function by incorporating a robust regularization term. This term leverages out-of-domain unlabeled data, guiding the classifier away from crowded, dense regions to improve overall performance and robustness.\n\n\u003cdiv align=\"center\"\u003e\n  \u003cimg src=\"https://raw.githubusercontent.com/deepmancer/rss-training-iclr2024/main/poster/images/pipeline.png\" alt=\"Overview of the RSS Training Framework\" style=\"max-width: 100%;\"\u003e\n  \u003cp\u003e\u003cstrong\u003eOverview of the RSS Training Framework\u003c/strong\u003e\u003c/p\u003e\n\u003c/div\u003e\n\n---\n\n## ✨ Key Contributions\n\nOur work introduces a polynomial-time framework that effectively integrates both labeled and slightly out-of-domain unlabeled data to improve generalization. Key contributions include:\n\n- **Non-asymptotic bounds** for both robust and non-robust learning scenarios.\n- **Enhanced generalization** over traditional ERM techniques, particularly when $n \\geq \\Omega(m^2/d)$.\n- **Dimension-independent** sample complexity under specific conditions.\n- **Improved sample complexity** from $O(d/\\epsilon^2)$ to $O(d/\\epsilon)$ when $n = O(d/\\epsilon^6)$.\n\n---\n\n## 📊 Conference Poster\n\n\u003cdiv align=\"center\"\u003e\n  \u003cimg src=\"https://raw.githubusercontent.com/deepmancer/rss-training-iclr2024/main/poster/poster.png\" style=\"max-width: 100%;\"\u003e\n  \u003cp\u003e\u003cstrong\u003eICLR 2024 Conference Poster\u003c/strong\u003e\u003c/p\u003e\n\u003c/div\u003e\n\nYou can access the LaTeX code for this poster [here](https://github.com/deepmancer/rss-training-iclr2024/tree/main/poster).\n\n---\n\n## 🚩 Updates\n\nWe are in the process of preparing the code for public release. Stay tuned for updates!\n\n---\n\n## 📚 Citation\n\nIf you find our work useful, please consider citing our paper:\n\n```bibtex\n@inproceedings{\nsaberi2024outofdomain,\ntitle={Out-Of-Domain Unlabeled Data Improves Generalization},\nauthor={Seyed Amir Hossein Saberi and Amir Najafi and Alireza Heidari and Mohammad Hosein Movasaghinia and Abolfazl Motahari and Babak Khalaj},\nbooktitle={The Twelfth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=Bo6GpQ3B9a}\n}\n```\n---\n\n## 📝 License\n\nThis project is licensed under the Apache 2.0 License. For more details, please see the [LICENSE](LICENSE) file.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fdeepmancer%2Frss-training-iclr2024","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fdeepmancer%2Frss-training-iclr2024","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fdeepmancer%2Frss-training-iclr2024/lists"}