{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "duchi-mcmahan-streeter-et-al-adaptive-online-gradients-2010", "title": "Adaptive online gradients: data-dependent proximal geometry and conditioning", "summary": "Developed data-dependent diagonal/full-matrix proximal or conditioning schemes from accumulated gradients, with complementary Adagrad and adaptive-bound formulations improving instance-dependent online regret.", "boundary": "Bundles independent2010McMahan/Streeter and2010preliminary/2011Duchi/Hazan/Singer contributions with source-specific author sets, not a fictitious five-author paper or identical algorithms. Competitive/regret statements restrict convex losses, feasible geometry, comparators and gradient assumptions; not global nonconvex deep-learning convergence or minimax improvement without structure.", "area": "Adaptive online convex optimization", "topic": "Artificial intelligence", "year": 2010, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "duchi-mcmahan-streeter-et-al-adaptive-online-gradients-2010", "target": { "record_id": "urn:uuid:7d8b2668-2216-416d-a158-83bdd552d111", "version": 1, "content_hash": "sha256:7113aff3a8eba812a22a7874233c432d74d9be47ba3b9cb2c1ba6dc1c6eefb88" }, "humans": [ "John Duchi", "Elad Hazan", "Yoram Singer", "H. Brendan McMahan", "Matthew Streeter" ], "authors_complete": true, "human_role_note": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261015-adagrad-original", "title": "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization", "url": "https://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf" }, { "id": "cur20261015-adaptive-bound-original", "title": "Adaptive Bound Optimization for Online Convex Optimization", "url": "https://arxiv.org/pdf/1002.4908v1" }, { "id": "cur20261015-online-conditioning-original", "title": "Less Regret via Online Conditioning", "url": "https://arxiv.org/pdf/1002.4862v1" } ], "points": "300000", "published_grade": { "id": "duchi-mcmahan-streeter-et-al-adaptive-online-gradients-2010", "entry_id": "duchi-mcmahan-streeter-et-al-adaptive-online-gradients-2010", "band": "S", "modifier": "-", "target": { "record_id": "urn:uuid:7d8b2668-2216-416d-a158-83bdd552d111", "version": 1, "content_hash": "sha256:7113aff3a8eba812a22a7874233c432d74d9be47ba3b9cb2c1ba6dc1c6eefb88" }, "what_changed": "Developed data-dependent diagonal/full-matrix proximal or conditioning schemes from accumulated gradients, with complementary Adagrad and adaptive-bound formulations improving instance-dependent online regret.", "scope": "Developed data-dependent diagonal/full-matrix proximal or conditioning schemes from accumulated gradients, with complementary Adagrad and adaptive-bound formulations improving instance-dependent online regret. Bundles independent2010McMahan/Streeter and2010preliminary/2011Duchi/Hazan/Singer contributions with source-specific author sets, not a fictitious five-author paper or identical algorithms. Competitive/regret statements restrict convex losses, feasible geometry, comparators and gradient assumptions; not global nonconvex deep-learning convergence or minimax improvement without structure.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:7d8b2668-2216-416d-a158-83bdd552d111", "version": 1, "content_hash": "sha256:7113aff3a8eba812a22a7874233c432d74d9be47ba3b9cb2c1ba6dc1c6eefb88" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-16T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "John Duchi", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization", "url": "https://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Elad Hazan", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization", "url": "https://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yoram Singer", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization", "url": "https://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "H. Brendan McMahan", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Adaptive Bound Optimization for Online Convex Optimization", "url": "https://arxiv.org/pdf/1002.4908v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Matthew Streeter", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Adaptive Bound Optimization for Online Convex Optimization", "url": "https://arxiv.org/pdf/1002.4908v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization", "url": "https://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf" } }, { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-1", "kind": "source_reference", "source": { "citation": "Adaptive Bound Optimization for Online Convex Optimization", "url": "https://arxiv.org/pdf/1002.4908v1" } }, { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-2", "kind": "source_reference", "source": { "citation": "Less Regret via Online Conditioning", "url": "https://arxiv.org/pdf/1002.4862v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "Bundles independent2010McMahan/Streeter and2010preliminary/2011Duchi/Hazan/Singer contributions with source-specific author sets, not a fictitious five-author paper or identical algorithms. Competitive/regret statements restrict convex losses, feasible geometry, comparators and gradient assumptions; not global nonconvex deep-learning convergence or minimax improvement without structure.", "domain": "Adaptive online convex optimization" }, "sources": [ { "citation": "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization", "url": "https://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf" }, { "citation": "Adaptive Bound Optimization for Online Convex Optimization", "url": "https://arxiv.org/pdf/1002.4908v1" }, { "citation": "Less Regret via Online Conditioning", "url": "https://arxiv.org/pdf/1002.4862v1" } ], "statement": "# Adaptive online gradients: data-dependent proximal geometry and conditioning\n\n## The question\n\nImagine adjusting many control knobs while learning from mistakes. Some knobs receive a strong signal almost every time; others are relevant only occasionally. Giving every knob the same step size can make the frequent ones move too aggressively while the rare ones barely get a chance. Adaptive gradient methods ask whether the past signals can guide future step sizes.\n\n## The idea\n\nA gradient describes how a small parameter change would affect the current loss, the number used to measure a mistake. The methods in this family accumulate information from earlier gradients and use it to change the geometry of later updates. In a familiar diagonal version, each coordinate keeps its own history. Directions with large accumulated squared gradients receive a smaller effective step.\n\n## Why it matters\n\nThis matters because the learning problem can reveal structure that a single fixed learning rate ignores. The source analyses connect such adaptation to regret: the accumulated loss compared with a specified reference choice. A regret bound is a statement under mathematical assumptions about losses, gradients and feasible choices. It does not promise that an arbitrary deep network reaches its globally best solution.\n\n## A teaching example\n\nConsider an original classroom illustration with two coordinates. Suppose their accumulated squared-gradient values are four and one, and ignore small stabilizing constants. Inverse-square-root scaling gives factors of one-half and one. The first coordinate therefore moves more cautiously. The calculation explains the adaptation rule; it is not an experiment showing that either coordinate is more scientifically important.\n\n## Scope and a useful next step\n\nImplement a tiny convex prediction task and log both the accumulated gradient history and effective update size. Compare with a carefully chosen fixed-step baseline. Keep data order and starting parameters identical. If adaptation helps only when features have unequal frequencies, that conditional finding explains more than an unsupported claim that the optimizer is always better. Bundles independent2010McMahan/Streeter and2010preliminary/2011Duchi/Hazan/Singer contributions with source-specific author sets, not a fictitious five-author paper or identical algorithms. Competitive/regret statements restrict convex losses, feasible geometry, comparators and gradient assumptions; not global nonconvex deep-learning convergence or minimax improvement without structure.\n\n## Scientific scope\n\nBundles independent2010McMahan/Streeter and2010preliminary/2011Duchi/Hazan/Singer contributions with source-specific author sets, not a fictitious five-author paper or identical algorithms. Competitive/regret statements restrict convex losses, feasible geometry, comparators and gradient assumptions; not global nonconvex deep-learning convergence or minimax improvement without structure.\n\n## Sources for this explanation\n\n- [Adaptive Subgradient Methods for Online Learning and Stochastic Optimization](https://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf): Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment.\n- [Adaptive Bound Optimization for Online Convex Optimization](https://arxiv.org/pdf/1002.4908v1): Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment.\n- [Less Regret via Online Conditioning](https://arxiv.org/pdf/1002.4862v1): Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment.\n", "summary": "Developed data-dependent diagonal/full-matrix proximal or conditioning schemes from accumulated gradients, with complementary Adagrad and adaptive-bound formulations improving instance-dependent online regret.", "title": "Adaptive online gradients: data-dependent proximal geometry and conditioning" }, "record_id": "urn:uuid:7d8b2668-2216-416d-a158-83bdd552d111", "version": 1, "source_ref": { "record_id": "urn:uuid:7d8b2668-2216-416d-a158-83bdd552d111", "version": 1, "content_hash": "sha256:7113aff3a8eba812a22a7874233c432d74d9be47ba3b9cb2c1ba6dc1c6eefb88" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:948e40ef386760802dbf9f7ce11dee0f09ca71746975722280ff5a8de5c008bd" }, "target": { "record_id": "urn:uuid:7d8b2668-2216-416d-a158-83bdd552d111", "version": 1, "content_hash": "sha256:7113aff3a8eba812a22a7874233c432d74d9be47ba3b9cb2c1ba6dc1c6eefb88" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "duchi-mcmahan-streeter-et-al-adaptive-online-gradients-2010", "entry_id": "duchi-mcmahan-streeter-et-al-adaptive-online-gradients-2010", "revision": 1, "title": "Adaptive online gradients: data-dependent proximal geometry and conditioning", "context": "Adaptive online convex optimization", "track": "Human-designed research method or resource", "band": "S", "modifier": "-", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-16", "evidence_cutoff": "2026-10-16", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "Developed data-dependent diagonal/full-matrix proximal or conditioning schemes from accumulated gradients, with complementary Adagrad and adaptive-bound formulations improving instance-dependent online regret.", "scope": "Developed data-dependent diagonal/full-matrix proximal or conditioning schemes from accumulated gradients, with complementary Adagrad and adaptive-bound formulations improving instance-dependent online regret. Bundles independent2010McMahan/Streeter and2010preliminary/2011Duchi/Hazan/Singer contributions with source-specific author sets, not a fictitious five-author paper or identical algorithms. Competitive/regret statements restrict convex losses, feasible geometry, comparators and gradient assumptions; not global nonconvex deep-learning convergence or minimax improvement without structure.", "evidence_summary": "The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Bundles independent2010McMahan/Streeter and2010preliminary/2011Duchi/Hazan/Singer contributions with source-specific author sets, not a fictitious five-author paper or identical algorithms. Competitive/regret statements restrict convex losses, feasible geometry, comparators and gradient assumptions; not global nonconvex deep-learning convergence or minimax improvement without structure.", "references": [ { "title": "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization", "url": "https://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." }, { "title": "Adaptive Bound Optimization for Online Convex Optimization", "url": "https://arxiv.org/pdf/1002.4908v1", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." }, { "title": "Less Regret via Online Conditioning", "url": "https://arxiv.org/pdf/1002.4862v1", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." } ], "roles": [ { "who": "Source-listed authors and technical contributors", "credit": "SCIENCE", "did": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities." } ], "target": { "record_id": "urn:uuid:7d8b2668-2216-416d-a158-83bdd552d111", "version": 1, "content_hash": "sha256:7113aff3a8eba812a22a7874233c432d74d9be47ba3b9cb2c1ba6dc1c6eefb88" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "300000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "duchi-mcmahan-streeter-et-al-adaptive-online-gradients-2010", "target": { "record_id": "urn:uuid:7d8b2668-2216-416d-a158-83bdd552d111", "version": 1, "content_hash": "sha256:7113aff3a8eba812a22a7874233c432d74d9be47ba3b9cb2c1ba6dc1c6eefb88" }, "sections": [ { "title": "The question", "text": "Imagine adjusting many control knobs while learning from mistakes. Some knobs receive a strong signal almost every time; others are relevant only occasionally. Giving every knob the same step size can make the frequent ones move too aggressively while the rare ones barely get a chance. Adaptive gradient methods ask whether the past signals can guide future step sizes." }, { "title": "The idea", "text": "A gradient describes how a small parameter change would affect the current loss, the number used to measure a mistake. The methods in this family accumulate information from earlier gradients and use it to change the geometry of later updates. In a familiar diagonal version, each coordinate keeps its own history. Directions with large accumulated squared gradients receive a smaller effective step." }, { "title": "Why it matters", "text": "This matters because the learning problem can reveal structure that a single fixed learning rate ignores. The source analyses connect such adaptation to regret: the accumulated loss compared with a specified reference choice. A regret bound is a statement under mathematical assumptions about losses, gradients and feasible choices. It does not promise that an arbitrary deep network reaches its globally best solution." }, { "title": "A teaching example", "text": "Consider an original classroom illustration with two coordinates. Suppose their accumulated squared-gradient values are four and one, and ignore small stabilizing constants. Inverse-square-root scaling gives factors of one-half and one. The first coordinate therefore moves more cautiously. The calculation explains the adaptation rule; it is not an experiment showing that either coordinate is more scientifically important." }, { "title": "Scope and a useful next step", "text": "Implement a tiny convex prediction task and log both the accumulated gradient history and effective update size. Compare with a carefully chosen fixed-step baseline. Keep data order and starting parameters identical. If adaptation helps only when features have unequal frequencies, that conditional finding explains more than an unsupported claim that the optimizer is always better. Bundles independent2010McMahan/Streeter and2010preliminary/2011Duchi/Hazan/Singer contributions with source-specific author sets, not a fictitious five-author paper or identical algorithms. Competitive/regret statements restrict convex losses, feasible geometry, comparators and gradient assumptions; not global nonconvex deep-learning convergence or minimax improvement without structure." } ], "references": [ { "title": "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization", "url": "https://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf", "locator": "Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment." }, { "title": "Adaptive Bound Optimization for Online Convex Optimization", "url": "https://arxiv.org/pdf/1002.4908v1", "locator": "Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment." }, { "title": "Less Regret via Online Conditioning", "url": "https://arxiv.org/pdf/1002.4862v1", "locator": "Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment." } ] }, "reading_status": "EXPLANATORY_GUIDE" }