{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "ba-kiros-hinton-layer-normalization-2016",
    "title": "LayerNorm: normalize each example without batch statistics",
    "summary": "Ba, Kiros and Hinton introduced layer normalization, computing activation mean and variance within each example instead of across a mini-batch, with learnable gains and biases and a recurrent-network formulation.\n\nBatch normalization uses statistics from other examples, which can complicate small-batch and recurrent training. LayerNorm instead computes its statistics from the hidden activations of the current example. The same normalization rule can run at training and test time, and recurrent networks can apply it separately at each step without storing a different set of batch statistics for every sequence position. The original 2016 paper reports useful training and generalization results across sequence, text-image, handwriting and digit tasks. It also reports that BatchNorm performs better in its preliminary convolutional-network experiments. LayerNorm changes the network computation, rather than merely rewriting the parameters of the same network. Its invariance analysis and practical benefits have conditions; it does not guarantee a global optimum or faster training for every architecture. The original three authors receive credit for this method, while later Transformer designs, implementations and refinements remain distinct.",
    "boundary": "Normalization is within one example, with learnable gain/bias and specified recurrent statistics. The stated invariances need their scaling/shift and nonzero-variance assumptions; they do not include arbitrary input translations or every per-unit transformation. Fixed numerical stabilizers can alter exact invariance. The experiments concern the reported architectures and training settings, not a universal performance or convergence guarantee.",
    "area": "Normalization for neural-network training",
    "topic": "Machine learning methods",
    "year": 2016,
    "lane": "HUMAN",
    "result": "ALGORITHM",
    "package_id": "ba-kiros-hinton-layer-normalization-2016",
    "target": {
      "record_id": "urn:uuid:85f219a3-23e7-4a88-a676-782811305d0f",
      "version": 1,
      "content_hash": "sha256:da5e7915609f7742d3a2eae3e887005b9921efba347ff44d21f89c6cd3706bcf"
    },
    "humans": [
      "Jimmy Lei Ba",
      "Jamie Ryan Kiros",
      "Geoffrey E. Hinton"
    ],
    "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"
    ],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "wrs24-layernorm2016-v1-original",
        "title": "Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1",
        "url": "https://arxiv.org/pdf/1607.06450v1"
      },
      {
        "id": "wrs24-layernorm2016-v1-original-metadata",
        "title": "Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1 repositoryidentity",
        "url": "https://arxiv.org/abs/1607.06450v1"
      }
    ],
    "points": "95000",
    "published_grade": {
      "id": "ba-kiros-hinton-layer-normalization-2016",
      "entry_id": "ba-kiros-hinton-layer-normalization-2016",
      "band": "A",
      "modifier": null,
      "target": {
        "record_id": "urn:uuid:85f219a3-23e7-4a88-a676-782811305d0f",
        "version": 1,
        "content_hash": "sha256:da5e7915609f7742d3a2eae3e887005b9921efba347ff44d21f89c6cd3706bcf"
      },
      "what_changed": "Ba, Kiros and Hinton introduced layer normalization, computing activation mean and variance within each example instead of across a mini-batch, with learnable gains and biases and a recurrent-network formulation.",
      "scope": "Ba, Kiros and Hinton introduced layer normalization, computing activation mean and variance within each example instead of across a mini-batch, with learnable gains and biases and a recurrent-network formulation. Normalization is within one example, with learnable gain/bias and specified recurrent statistics. The stated invariances need their scaling/shift and nonzero-variance assumptions; they do not include arbitrary input translations or every per-unit transformation. Fixed numerical stabilizers can alter exact invariance. The experiments concern the reported architectures and training settings, not a universal performance or convergence guarantee.",
      "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:85f219a3-23e7-4a88-a676-782811305d0f",
        "version": 1,
        "content_hash": "sha256:da5e7915609f7742d3a2eae3e887005b9921efba347ff44d21f89c6cd3706bcf"
      },
      "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": "Jimmy Lei Ba",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1",
            "url": "https://arxiv.org/pdf/1607.06450v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Jamie Ryan Kiros",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1",
            "url": "https://arxiv.org/pdf/1607.06450v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Geoffrey E. Hinton",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1",
            "url": "https://arxiv.org/pdf/1607.06450v1"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1",
            "url": "https://arxiv.org/pdf/1607.06450v1"
          }
        },
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-1",
          "kind": "source_reference",
          "source": {
            "citation": "Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1 repositoryidentity",
            "url": "https://arxiv.org/abs/1607.06450v1"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "Normalization is within one example, with learnable gain/bias and specified recurrent statistics. The stated invariances need their scaling/shift and nonzero-variance assumptions; they do not include arbitrary input translations or every per-unit transformation. Fixed numerical stabilizers can alter exact invariance. The experiments concern the reported architectures and training settings, not a universal performance or convergence guarantee.",
        "domain": "Normalization for neural-network training"
      },
      "sources": [
        {
          "citation": "Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1",
          "url": "https://arxiv.org/pdf/1607.06450v1"
        },
        {
          "citation": "Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1 repositoryidentity",
          "url": "https://arxiv.org/abs/1607.06450v1"
        }
      ],
      "statement": "# LayerNorm: normalize each example without batch statistics\n\nBa, Kiros and Hinton introduced layer normalization, computing activation mean and variance within each example instead of across a mini-batch, with learnable gains and biases and a recurrent-network formulation.\n\nBatch normalization uses statistics from other examples, which can complicate small-batch and recurrent training. LayerNorm instead computes its statistics from the hidden activations of the current example. The same normalization rule can run at training and test time, and recurrent networks can apply it separately at each step without storing a different set of batch statistics for every sequence position. The original 2016 paper reports useful training and generalization results across sequence, text-image, handwriting and digit tasks. It also reports that BatchNorm performs better in its preliminary convolutional-network experiments. LayerNorm changes the network computation, rather than merely rewriting the parameters of the same network. Its invariance analysis and practical benefits have conditions; it does not guarantee a global optimum or faster training for every architecture. The original three authors receive credit for this method, while later Transformer designs, implementations and refinements remain distinct.\n\n## Scientific scope\n\nNormalization is within one example, with learnable gain/bias and specified recurrent statistics. The stated invariances need their scaling/shift and nonzero-variance assumptions; they do not include arbitrary input translations or every per-unit transformation. Fixed numerical stabilizers can alter exact invariance. The experiments concern the reported architectures and training settings, not a universal performance or convergence guarantee.\n\n## People and systems\n\nJimmy Lei Ba; Jamie Ryan Kiros; Geoffrey E. Hinton\n\n\n\n## Original sources\n\n- [Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1](https://arxiv.org/pdf/1607.06450v1)\n- [Ba,KirosandHinton,LayerNormalization,exact2016arXiv1607.06450v1 repositoryidentity](https://arxiv.org/abs/1607.06450v1)\n",
      "summary": "Ba, Kiros and Hinton introduced layer normalization, computing activation mean and variance within each example instead of across a mini-batch, with learnable gains and biases and a recurrent-network formulation.\n\nBatch normalization uses statistics from other examples, which can complicate small-batch and recurrent training. LayerNorm instead computes its statistics from the hidden activations of the current example. The same normalization rule can run at training and test time, and recurrent networks can apply it separately at each step without storing a different set of batch statistics for every sequence position. The original 2016 paper reports useful training and generalization results across sequence, text-image, handwriting and digit tasks. It also reports that BatchNorm performs better in its preliminary convolutional-network experiments. LayerNorm changes the network computation, rather than merely rewriting the parameters of the same network. Its invariance analysis and practical benefits have conditions; it does not guarantee a global optimum or faster training for every architecture. The original three authors receive credit for this method, while later Transformer designs, implementations and refinements remain distinct.",
      "title": "LayerNorm: normalize each example without batch statistics"
    },
    "record_id": "urn:uuid:85f219a3-23e7-4a88-a676-782811305d0f",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:85f219a3-23e7-4a88-a676-782811305d0f",
      "version": 1,
      "content_hash": "sha256:da5e7915609f7742d3a2eae3e887005b9921efba347ff44d21f89c6cd3706bcf"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:5c0b525eadf58d2c929467441391fdf46d54f0ea714494f0774d1bade0f4e7bf"
  },
  "target": {
    "record_id": "urn:uuid:85f219a3-23e7-4a88-a676-782811305d0f",
    "version": 1,
    "content_hash": "sha256:da5e7915609f7742d3a2eae3e887005b9921efba347ff44d21f89c6cd3706bcf"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "ba-kiros-hinton-layer-normalization-2016",
    "entry_id": "ba-kiros-hinton-layer-normalization-2016",
    "revision": 1,
    "title": "LayerNorm: normalize each example without batch statistics",
    "context": "Normalization for neural-network training",
    "track": "Normalization for neural-network training",
    "band": "A",
    "modifier": null,
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-04",
    "evidence_cutoff": "2026-10-04",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "Ba, Kiros and Hinton introduced layer normalization, computing activation mean and variance within each example instead of across a mini-batch, with learnable gains and biases and a recurrent-network formulation.",
    "scope": "Ba, Kiros and Hinton introduced layer normalization, computing activation mean and variance within each example instead of across a mini-batch, with learnable gains and biases and a recurrent-network formulation. Normalization is within one example, with learnable gain/bias and specified recurrent statistics. The stated invariances need their scaling/shift and nonzero-variance assumptions; they do not include arbitrary input translations or every per-unit transformation. Fixed numerical stabilizers can alter exact invariance. The experiments concern the reported architectures and training settings, not a universal performance or convergence guarantee.",
    "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": "Ba, Kirosand Hinton, Layer Normalization, exact 2016 ar Xiv 1607.06450 v 1",
        "url": "https://arxiv.org/pdf/1607.06450v1",
        "locator": "Original linked work. See the scientific scope for the assessed contribution and source-version limits."
      },
      {
        "title": "Ba, Kirosand Hinton, Layer Normalization, exact 2016 ar Xiv 1607.06450 v 1 repositoryidentity",
        "url": "https://arxiv.org/abs/1607.06450v1",
        "locator": "Original linked work. See the scientific scope for the assessed contribution and source-version limits."
      }
    ],
    "roles": [
      {
        "who": "Jimmy Lei Ba",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      },
      {
        "who": "Jamie Ryan Kiros",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      },
      {
        "who": "Geoffrey E. Hinton",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      }
    ],
    "target": {
      "record_id": "urn:uuid:85f219a3-23e7-4a88-a676-782811305d0f",
      "version": 1,
      "content_hash": "sha256:da5e7915609f7742d3a2eae3e887005b9921efba347ff44d21f89c6cd3706bcf"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "95000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": null,
  "reading_status": "SOURCE_LINKED_NOTE"
}