{"id":29830827,"url":"https://github.com/finite-sample/pareto-gd","last_synced_at":"2026-02-09T04:04:47.390Z","repository":{"id":301639789,"uuid":"1009893119","full_name":"finite-sample/pareto-gd","owner":"finite-sample","description":"Pareto Improving GD. Only update when when doing so improves or maintains performance on all previously seen data.","archived":false,"fork":false,"pushed_at":"2025-07-16T20:38:02.000Z","size":1316,"stargazers_count":1,"open_issues_count":0,"forks_count":0,"subscribers_count":1,"default_branch":"main","last_synced_at":"2025-09-26T23:38:28.479Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":null,"language":"Jupyter Notebook","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/finite-sample.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"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,"zenodo":null}},"created_at":"2025-06-27T22:53:45.000Z","updated_at":"2025-07-16T20:38:06.000Z","dependencies_parsed_at":"2025-06-27T23:30:59.841Z","dependency_job_id":"ea05f65c-702a-49c3-984f-772f349b6889","html_url":"https://github.com/finite-sample/pareto-gd","commit_stats":null,"previous_names":["finite-sample/pareto-gd"],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/finite-sample/pareto-gd","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/finite-sample%2Fpareto-gd","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/finite-sample%2Fpareto-gd/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/finite-sample%2Fpareto-gd/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/finite-sample%2Fpareto-gd/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/finite-sample","download_url":"https://codeload.github.com/finite-sample/pareto-gd/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/finite-sample%2Fpareto-gd/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":29255980,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-02-09T03:07:45.136Z","status":"ssl_error","status_checked_at":"2026-02-09T03:07:24.123Z","response_time":56,"last_error":"SSL_connect returned=1 errno=0 peeraddr=140.82.121.6:443 state=error: unexpected eof while reading","robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":false,"can_crawl_api":true,"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":[],"created_at":"2025-07-29T10:11:40.954Z","updated_at":"2026-02-09T04:04:47.378Z","avatar_url":"https://github.com/finite-sample.png","language":"Jupyter Notebook","funding_links":[],"categories":[],"sub_categories":[],"readme":"# (Don't) Forget About It: Forgetting-Penalized Supervised Learning\n\n## Background\n\nA well-known problem in machine learning is **regression**: as models update, they sometimes \"forget\" how to correctly handle examples they previously got right. This is especially frustrating in production or user-facing systems, where a model suddenly failing on known-good cases can be more disruptive than missing new ones.\n\nCatastrophic forgetting is well-studied in **continual learning** (French, 1999), and rehearsal/buffer methods are common. But for standard supervised learning, less attention has been paid to *actively penalizing regression* during ordinary training.\n\n---\n\n## Approaches Compared\n\nWe compare three strategies:\n\n**1. Standard Training (Baseline)**  \nThe usual approach—minimize training loss with no explicit mechanism to prevent forgetting.\n\n**2. Forgetting-Penalized Training**  \nInspired by continual learning methods like Elastic Weight Consolidation (Kirkpatrick et al., 2017), this adds a penalty whenever an example previously classified correctly becomes incorrect. It discourages \"unlearning,\" but does not eliminate all changes.\n\n**3. Soft Pareto-Penalized Training**  \nDrawing on Pareto-improvement ideas and recent multi-task optimization research (Lin et al., 2019; Navon et al., 2021), this method penalizes *any* increase in per-example loss—not just flips from correct to incorrect. It enforces a softer, broader \"do no harm\" principle across all training examples.\n\n---\n\n## Experiment\n\nOn the Adult income dataset, we trained all three methods with identical neural network architectures. Penalties were introduced after a warmup period, allowing the model to stabilize before beginning to penalize regressions.\n\n---\n\n## Results\n\n| Method           | Total Forgetting | Final Train Acc | Final Val Acc |\n|------------------|------------------|-----------------|---------------|\n| **Baseline**     | 5668             | 0.794           | 0.788         |\n| **Forgetting Pen.** | 122          | 0.759           | 0.760         |\n| **Soft Pareto**  | 290              | 0.786           | 0.783         |\n\n- Both penalized methods reduced forgetting by an order of magnitude **compared to baseline**.\n- **Soft Pareto** provided a strong trade-off: low forgetting with minimal accuracy loss.\n- **Forgetting-Penalized** achieved the lowest forgetting, but at a more significant cost to accuracy.\n- **Baseline training** delivered the highest accuracy—but experienced frequent regression.\n\n---\n\n## Contribution\n\nWhile regularization and continual learning are well-established, our work shows that **simple, lightweight penalty-based mechanisms**—added to ordinary training—can greatly reduce regression *without substantial accuracy loss*. The **Soft Pareto loss** is especially practical, implementing a “do no harm” bias that’s easy to integrate.\n\n---\n\n## Where It Matters\n\n- **Production-grade systems** where regression on known-good cases is unacceptable.\n- **Human-facing models** where consistency matters to user trust.\n- **High-stakes domains** like medical, fraud detection, or compliance.\n- **Curriculum or staged learning setups**, where early learning shouldn't be overwritten by later stages.\n\n---\n\n## Summary\n\nIf maintaining correctness on previously learned examples matters—even under normal supervised training—then adding **penalty terms** for forgetting or loss regression is effective, easy to implement, and provides a natural “Pareto bias” in practice.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ffinite-sample%2Fpareto-gd","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Ffinite-sample%2Fpareto-gd","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ffinite-sample%2Fpareto-gd/lists"}