{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "wu-he-group-normalization-2018",
    "title": "GroupNorm: train visual models with small batches",
    "summary": "Wu and He introduced group normalization, computing mean and variance across spatial positions and groups of channels within each example, enabling visual-network normalization without batch statistics.\n\nLarge images and video clips can leave room for only a few training examples on each GPU. GroupNorm helps avoid relying on statistics from that small batch: it splits feature channels into groups and normalizes each group within each example. Unlike a group convolution, it can be inserted into an ordinary network without requiring grouped convolutional layers. In the reported ResNet-50 ImageNet experiment with two images per GPU, error falls from 34.7% with unsynchronized BatchNorm to 24.1% with GroupNorm. With 32 images per GPU, BatchNorm instead has the lower error. The paper also studies object detection, segmentation and video classification under stated training settings. Group count remains a design choice, and freedom from batch statistics does not mean optimization is unaffected by batch size. The 2018 revised preprint keeps its own result version; later code, models and applications are not automatically included in the original contribution.",
    "boundary": "Normalization groups channels within each example under a chosen group count and architecture. It avoids batch-statistics dependence but does not give a universal invariance, convergence or quality guarantee. Extreme group counts connect to different normalization cases, and empirical results depend on task/training settings. WRS has not independently replayed the benchmarks.",
    "area": "Normalization for neural-network training",
    "topic": "Machine learning methods",
    "year": 2018,
    "lane": "HUMAN",
    "result": "ALGORITHM",
    "package_id": "wu-he-group-normalization-2018",
    "target": {
      "record_id": "urn:uuid:756f1a61-b85c-4edf-a861-02f0e7bce28f",
      "version": 1,
      "content_hash": "sha256:fdb0775d50b40bebfd97d36f205433a84a232067b16af6efecac1243533a22a7"
    },
    "humans": [
      "Yuxin Wu",
      "Kaiming He"
    ],
    "authors_complete": true,
    "human_role_note": "Source-listed contributors to this scoped work. Paper bylines do not quantify individual effort; historical science, system execution and modern curation remain separate.",
    "agents": [],
    "related": [
      "ioffe-szegedy-batch-normalization-2015",
      "ba-kiros-hinton-layer-normalization-2016"
    ],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "wrs24-groupnorm2018-v3-original",
        "title": "WuandHe,GroupNormalization,exact2018arXiv1803.08494v3",
        "url": "https://arxiv.org/pdf/1803.08494v3"
      },
      {
        "id": "wrs24-groupnorm2018-v3-original-metadata",
        "title": "WuandHe,GroupNormalization,exact2018arXiv1803.08494v3 repositoryidentity",
        "url": "https://arxiv.org/abs/1803.08494v3"
      }
    ],
    "points": "45000",
    "published_grade": {
      "id": "wu-he-group-normalization-2018",
      "entry_id": "wu-he-group-normalization-2018",
      "band": "B",
      "modifier": null,
      "target": {
        "record_id": "urn:uuid:756f1a61-b85c-4edf-a861-02f0e7bce28f",
        "version": 1,
        "content_hash": "sha256:fdb0775d50b40bebfd97d36f205433a84a232067b16af6efecac1243533a22a7"
      },
      "what_changed": "Wu and He introduced group normalization, computing mean and variance across spatial positions and groups of channels within each example, enabling visual-network normalization without batch statistics.",
      "scope": "Wu and He introduced group normalization, computing mean and variance across spatial positions and groups of channels within each example, enabling visual-network normalization without batch statistics. Normalization groups channels within each example under a chosen group count and architecture. It avoids batch-statistics dependence but does not give a universal invariance, convergence or quality guarantee. Extreme group counts connect to different normalization cases, and empirical results depend on task/training settings. WRS has not independently replayed the benchmarks.",
      "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:756f1a61-b85c-4edf-a861-02f0e7bce28f",
        "version": 1,
        "content_hash": "sha256:fdb0775d50b40bebfd97d36f205433a84a232067b16af6efecac1243533a22a7"
      },
      "edition": "ratings-e039106f30dec177"
    }
  },
  "version": {
    "kind": "PUBLIC_RECORD_VIEW",
    "previous": null,
    "published_at": "2026-10-03T00:00:00Z",
    "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": "Yuxin Wu",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "WuandHe,GroupNormalization,exact2018arXiv1803.08494v3",
            "url": "https://arxiv.org/pdf/1803.08494v3"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Kaiming He",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "WuandHe,GroupNormalization,exact2018arXiv1803.08494v3",
            "url": "https://arxiv.org/pdf/1803.08494v3"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "WuandHe,GroupNormalization,exact2018arXiv1803.08494v3",
            "url": "https://arxiv.org/pdf/1803.08494v3"
          }
        },
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-1",
          "kind": "source_reference",
          "source": {
            "citation": "WuandHe,GroupNormalization,exact2018arXiv1803.08494v3 repositoryidentity",
            "url": "https://arxiv.org/abs/1803.08494v3"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "Normalization groups channels within each example under a chosen group count and architecture. It avoids batch-statistics dependence but does not give a universal invariance, convergence or quality guarantee. Extreme group counts connect to different normalization cases, and empirical results depend on task/training settings. WRS has not independently replayed the benchmarks.",
        "domain": "Normalization for neural-network training"
      },
      "sources": [
        {
          "citation": "WuandHe,GroupNormalization,exact2018arXiv1803.08494v3",
          "url": "https://arxiv.org/pdf/1803.08494v3"
        },
        {
          "citation": "WuandHe,GroupNormalization,exact2018arXiv1803.08494v3 repositoryidentity",
          "url": "https://arxiv.org/abs/1803.08494v3"
        }
      ],
      "statement": "# GroupNorm: train visual models with small batches\n\nWu and He introduced group normalization, computing mean and variance across spatial positions and groups of channels within each example, enabling visual-network normalization without batch statistics.\n\nLarge images and video clips can leave room for only a few training examples on each GPU. GroupNorm helps avoid relying on statistics from that small batch: it splits feature channels into groups and normalizes each group within each example. Unlike a group convolution, it can be inserted into an ordinary network without requiring grouped convolutional layers. In the reported ResNet-50 ImageNet experiment with two images per GPU, error falls from 34.7% with unsynchronized BatchNorm to 24.1% with GroupNorm. With 32 images per GPU, BatchNorm instead has the lower error. The paper also studies object detection, segmentation and video classification under stated training settings. Group count remains a design choice, and freedom from batch statistics does not mean optimization is unaffected by batch size. The 2018 revised preprint keeps its own result version; later code, models and applications are not automatically included in the original contribution.\n\n## Scientific scope\n\nNormalization groups channels within each example under a chosen group count and architecture. It avoids batch-statistics dependence but does not give a universal invariance, convergence or quality guarantee. Extreme group counts connect to different normalization cases, and empirical results depend on task/training settings. WRS has not independently replayed the benchmarks.\n\n## People and systems\n\nYuxin Wu; Kaiming He\n\n\n\n## Original sources\n\n- [WuandHe,GroupNormalization,exact2018arXiv1803.08494v3](https://arxiv.org/pdf/1803.08494v3)\n- [WuandHe,GroupNormalization,exact2018arXiv1803.08494v3 repositoryidentity](https://arxiv.org/abs/1803.08494v3)\n",
      "summary": "Wu and He introduced group normalization, computing mean and variance across spatial positions and groups of channels within each example, enabling visual-network normalization without batch statistics.\n\nLarge images and video clips can leave room for only a few training examples on each GPU. GroupNorm helps avoid relying on statistics from that small batch: it splits feature channels into groups and normalizes each group within each example. Unlike a group convolution, it can be inserted into an ordinary network without requiring grouped convolutional layers. In the reported ResNet-50 ImageNet experiment with two images per GPU, error falls from 34.7% with unsynchronized BatchNorm to 24.1% with GroupNorm. With 32 images per GPU, BatchNorm instead has the lower error. The paper also studies object detection, segmentation and video classification under stated training settings. Group count remains a design choice, and freedom from batch statistics does not mean optimization is unaffected by batch size. The 2018 revised preprint keeps its own result version; later code, models and applications are not automatically included in the original contribution.",
      "title": "GroupNorm: train visual models with small batches"
    },
    "record_id": "urn:uuid:756f1a61-b85c-4edf-a861-02f0e7bce28f",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:756f1a61-b85c-4edf-a861-02f0e7bce28f",
      "version": 1,
      "content_hash": "sha256:fdb0775d50b40bebfd97d36f205433a84a232067b16af6efecac1243533a22a7"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:e2e38619ac78bdc8b67480a14b1fc9312074bd82171080cb15e5544aa56c31fb"
  },
  "target": {
    "record_id": "urn:uuid:756f1a61-b85c-4edf-a861-02f0e7bce28f",
    "version": 1,
    "content_hash": "sha256:fdb0775d50b40bebfd97d36f205433a84a232067b16af6efecac1243533a22a7"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "wu-he-group-normalization-2018",
    "entry_id": "wu-he-group-normalization-2018",
    "revision": 1,
    "title": "GroupNorm: train visual models with small batches",
    "context": "Normalization for neural-network training",
    "track": "Normalization for neural-network training",
    "band": "B",
    "modifier": null,
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-04",
    "evidence_cutoff": "2026-10-04",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "Wu and He introduced group normalization, computing mean and variance across spatial positions and groups of channels within each example, enabling visual-network normalization without batch statistics.",
    "scope": "Wu and He introduced group normalization, computing mean and variance across spatial positions and groups of channels within each example, enabling visual-network normalization without batch statistics. Normalization groups channels within each example under a chosen group count and architecture. It avoids batch-statistics dependence but does not give a universal invariance, convergence or quality guarantee. Extreme group counts connect to different normalization cases, and empirical results depend on task/training settings. WRS has not independently replayed the benchmarks.",
    "evidence_summary": "Source-linked, scoped account of the original work. The published scope states the scientific assumptions and remaining verification limits. WRS has not independently certified the full proof, experimental result or system benchmark.",
    "references": [
      {
        "title": "Wuand He, Group Normalization, exact 2018 ar Xiv 1803.08494 v 3",
        "url": "https://arxiv.org/pdf/1803.08494v3",
        "locator": "Original linked work. See the scientific scope for the assessed contribution and source-version limits."
      },
      {
        "title": "Wuand He, Group Normalization, exact 2018 ar Xiv 1803.08494 v 3 repositoryidentity",
        "url": "https://arxiv.org/abs/1803.08494v3",
        "locator": "Original linked work. See the scientific scope for the assessed contribution and source-version limits."
      }
    ],
    "roles": [
      {
        "who": "Yuxin Wu",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      },
      {
        "who": "Kaiming He",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      }
    ],
    "target": {
      "record_id": "urn:uuid:756f1a61-b85c-4edf-a861-02f0e7bce28f",
      "version": 1,
      "content_hash": "sha256:fdb0775d50b40bebfd97d36f205433a84a232067b16af6efecac1243533a22a7"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "45000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": null,
  "reading_status": "SOURCE_LINKED_NOTE"
}