{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2021:Z3QZVUI6UTIZR5M5HYQSJDLIKY","short_pith_number":"pith:Z3QZVUI6","canonical_record":{"source":{"id":"2105.04906","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2021-05-11T09:53:21Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"8bc0a69c90c72dd6963df5adfd4d9898ef7492bc826dadfa623cd2f91e796c9a","abstract_canon_sha256":"369d5bbba7f095e3dc758d10cbbf40559d032349b243141452acb338d3ae81e9"},"schema_version":"1.0"},"canonical_sha256":"cee19ad11ea4d198f59d3e21248d685611db12ed366cbcf1f061fe9bdcde2786","source":{"kind":"arxiv","id":"2105.04906","version":3},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2105.04906","created_at":"2026-05-17T23:38:51Z"},{"alias_kind":"arxiv_version","alias_value":"2105.04906v3","created_at":"2026-05-17T23:38:51Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2105.04906","created_at":"2026-05-17T23:38:51Z"},{"alias_kind":"pith_short_12","alias_value":"Z3QZVUI6UTIZ","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"Z3QZVUI6UTIZR5M5","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"Z3QZVUI6","created_at":"2026-05-18T12:33:33Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2021:Z3QZVUI6UTIZR5M5HYQSJDLIKY","target":"record","payload":{"canonical_record":{"source":{"id":"2105.04906","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2021-05-11T09:53:21Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"8bc0a69c90c72dd6963df5adfd4d9898ef7492bc826dadfa623cd2f91e796c9a","abstract_canon_sha256":"369d5bbba7f095e3dc758d10cbbf40559d032349b243141452acb338d3ae81e9"},"schema_version":"1.0"},"canonical_sha256":"cee19ad11ea4d198f59d3e21248d685611db12ed366cbcf1f061fe9bdcde2786","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:51.093498Z","signature_b64":"+yZq5KRN4VoWk0UkqRRGThxSF4dM8vxyPrhQPU8akyfmfaxML1zTZww38ebx40VW7dml97LYKbDh+Qgq1CTTBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cee19ad11ea4d198f59d3e21248d685611db12ed366cbcf1f061fe9bdcde2786","last_reissued_at":"2026-05-17T23:38:51.093051Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:51.093051Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2105.04906","source_version":3,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:51Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"B8KPAH9NEXagnbgsywmBgcCVP54cRZPKELXfYxN1/O8yfBAS9lasYdMgDlnCDiEoMDsET71MQtPb7m9wUrVSBA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-31T16:28:35.036966Z"},"content_sha256":"5d735262715b61a0d1fc368293de486e01122c4c7bf470e2c67c7927762f8738","schema_version":"1.0","event_id":"sha256:5d735262715b61a0d1fc368293de486e01122c4c7bf470e2c67c7927762f8738"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2021:Z3QZVUI6UTIZR5M5HYQSJDLIKY","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"VICReg prevents collapse to constant embeddings in self-supervised learning by adding an explicit variance term per dimension plus covariance regularization.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CV","authors_text":"Adrien Bardes, Jean Ponce, Yann LeCun","submitted_at":"2021-05-11T09:53:21Z","abstract_excerpt":"Recent self-supervised methods for image representation learning are based on maximizing the agreement between embedding vectors from different views of the same image. A trivial solution is obtained when the encoder outputs constant vectors. This collapse problem is often avoided through implicit biases in the learning architecture, that often lack a clear justification or interpretation. In this paper, we introduce VICReg (Variance-Invariance-Covariance Regularization), a method that explicitly avoids the collapse problem with a simple regularization term on the variance of the embeddings al"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"VICReg achieves results on par with the state of the art on several downstream tasks. In addition, we show that incorporating our new variance term into other methods helps stabilize the training and leads to performance improvements.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that enforcing per-dimension variance above a fixed threshold (combined with covariance regularization) is sufficient to eliminate collapse across architectures and datasets without introducing new failure modes or requiring architecture-specific adjustments.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"VICReg prevents collapse in self-supervised image embeddings via explicit variance, invariance, and covariance regularization and matches state-of-the-art downstream performance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"VICReg prevents collapse to constant embeddings in self-supervised learning by adding an explicit variance term per dimension plus covariance regularization.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e13e9144756b14cdc9363998f526827094b193b882c98f501206225fcd0dd7a7"},"source":{"id":"2105.04906","kind":"arxiv","version":3},"verdict":{"id":"59a6ffca-94b0-439a-be4e-b6a9f49cc80a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T15:25:02.646360Z","strongest_claim":"VICReg achieves results on par with the state of the art on several downstream tasks. In addition, we show that incorporating our new variance term into other methods helps stabilize the training and leads to performance improvements.","one_line_summary":"VICReg prevents collapse in self-supervised image embeddings via explicit variance, invariance, and covariance regularization and matches state-of-the-art downstream performance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that enforcing per-dimension variance above a fixed threshold (combined with covariance regularization) is sufficient to eliminate collapse across architectures and datasets without introducing new failure modes or requiring architecture-specific adjustments.","pith_extraction_headline":"VICReg prevents collapse to constant embeddings in self-supervised learning by adding an explicit variance term per dimension plus covariance regularization."},"references":{"count":119,"sample":[{"doi":"","year":2020,"title":"Self-labelling via simultaneous clustering and representation learning","work_id":"ee3a6975-6d06-42fc-a871-fc46e4b378df","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Learning representations by maximizing mutual information across views","work_id":"5d83c1d1-dde5-4d4c-87b5-3e817a94e2b8","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"Bautista, Artsiom Sanakoyeu, Ekaterina Sutter, and Björn Ommer","work_id":"dcc4f8bb-644e-4890-9aad-7988354f39dc","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1994,"title":"Signature verification using a “siamese” time delay neural network","work_id":"bb19cefe-4252-47d8-898b-d48d1ed8d393","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Deep clustering for unsupervised learning","work_id":"0ed1fc50-5600-4155-ba9b-160345bb7140","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":119,"snapshot_sha256":"a5f17ceafdd0901618c1e3233abfbf0db44bd44c9bc9e6b14918a506cd73ba64","internal_anchors":5},"formal_canon":{"evidence_count":3,"snapshot_sha256":"41b1d7d90914900408d3d2a21e92974a5524d8b0b60fc48009463740ad93a0c6"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"59a6ffca-94b0-439a-be4e-b6a9f49cc80a"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:51Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"MlpQtr5sed7CDXICfn5LLW6jiu0tUg3ruLFVCl+IK8TVSefKjC7Y9BNcWFgQzewAjTOCLDPrdKXg8J2fZI05Cg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-31T16:28:35.037665Z"},"content_sha256":"aeb997beb8c2e25fcfde825cdd6fcf0fae2cbf682126c20ce4847396aa6ac9b7","schema_version":"1.0","event_id":"sha256:aeb997beb8c2e25fcfde825cdd6fcf0fae2cbf682126c20ce4847396aa6ac9b7"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/Z3QZVUI6UTIZR5M5HYQSJDLIKY/bundle.json","state_url":"https://pith.science/pith/Z3QZVUI6UTIZR5M5HYQSJDLIKY/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/Z3QZVUI6UTIZR5M5HYQSJDLIKY/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-31T16:28:35Z","links":{"resolver":"https://pith.science/pith/Z3QZVUI6UTIZR5M5HYQSJDLIKY","bundle":"https://pith.science/pith/Z3QZVUI6UTIZR5M5HYQSJDLIKY/bundle.json","state":"https://pith.science/pith/Z3QZVUI6UTIZR5M5HYQSJDLIKY/state.json","well_known_bundle":"https://pith.science/.well-known/pith/Z3QZVUI6UTIZR5M5HYQSJDLIKY/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2021:Z3QZVUI6UTIZR5M5HYQSJDLIKY","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"369d5bbba7f095e3dc758d10cbbf40559d032349b243141452acb338d3ae81e9","cross_cats_sorted":["cs.AI","cs.LG"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2021-05-11T09:53:21Z","title_canon_sha256":"8bc0a69c90c72dd6963df5adfd4d9898ef7492bc826dadfa623cd2f91e796c9a"},"schema_version":"1.0","source":{"id":"2105.04906","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2105.04906","created_at":"2026-05-17T23:38:51Z"},{"alias_kind":"arxiv_version","alias_value":"2105.04906v3","created_at":"2026-05-17T23:38:51Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2105.04906","created_at":"2026-05-17T23:38:51Z"},{"alias_kind":"pith_short_12","alias_value":"Z3QZVUI6UTIZ","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"Z3QZVUI6UTIZR5M5","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"Z3QZVUI6","created_at":"2026-05-18T12:33:33Z"}],"graph_snapshots":[{"event_id":"sha256:aeb997beb8c2e25fcfde825cdd6fcf0fae2cbf682126c20ce4847396aa6ac9b7","target":"graph","created_at":"2026-05-17T23:38:51Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"VICReg achieves results on par with the state of the art on several downstream tasks. In addition, we show that incorporating our new variance term into other methods helps stabilize the training and leads to performance improvements."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"The assumption that enforcing per-dimension variance above a fixed threshold (combined with covariance regularization) is sufficient to eliminate collapse across architectures and datasets without introducing new failure modes or requiring architecture-specific adjustments."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"VICReg prevents collapse in self-supervised image embeddings via explicit variance, invariance, and covariance regularization and matches state-of-the-art downstream performance."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"VICReg prevents collapse to constant embeddings in self-supervised learning by adding an explicit variance term per dimension plus covariance regularization."}],"snapshot_sha256":"e13e9144756b14cdc9363998f526827094b193b882c98f501206225fcd0dd7a7"},"formal_canon":{"evidence_count":3,"snapshot_sha256":"41b1d7d90914900408d3d2a21e92974a5524d8b0b60fc48009463740ad93a0c6"},"paper":{"abstract_excerpt":"Recent self-supervised methods for image representation learning are based on maximizing the agreement between embedding vectors from different views of the same image. A trivial solution is obtained when the encoder outputs constant vectors. This collapse problem is often avoided through implicit biases in the learning architecture, that often lack a clear justification or interpretation. In this paper, we introduce VICReg (Variance-Invariance-Covariance Regularization), a method that explicitly avoids the collapse problem with a simple regularization term on the variance of the embeddings al","authors_text":"Adrien Bardes, Jean Ponce, Yann LeCun","cross_cats":["cs.AI","cs.LG"],"headline":"VICReg prevents collapse to constant embeddings in self-supervised learning by adding an explicit variance term per dimension plus covariance regularization.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2021-05-11T09:53:21Z","title":"VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning"},"references":{"count":119,"internal_anchors":5,"resolved_work":119,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Self-labelling via simultaneous clustering and representation learning","work_id":"ee3a6975-6d06-42fc-a871-fc46e4b378df","year":2020},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Learning representations by maximizing mutual information across views","work_id":"5d83c1d1-dde5-4d4c-87b5-3e817a94e2b8","year":2019},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Bautista, Artsiom Sanakoyeu, Ekaterina Sutter, and Björn Ommer","work_id":"dcc4f8bb-644e-4890-9aad-7988354f39dc","year":2016},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Signature verification using a “siamese” time delay neural network","work_id":"bb19cefe-4252-47d8-898b-d48d1ed8d393","year":1994},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Deep clustering for unsupervised learning","work_id":"0ed1fc50-5600-4155-ba9b-160345bb7140","year":2018}],"snapshot_sha256":"a5f17ceafdd0901618c1e3233abfbf0db44bd44c9bc9e6b14918a506cd73ba64"},"source":{"id":"2105.04906","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-15T15:25:02.646360Z","id":"59a6ffca-94b0-439a-be4e-b6a9f49cc80a","model_set":{"reader":"grok-4.3"},"one_line_summary":"VICReg prevents collapse in self-supervised image embeddings via explicit variance, invariance, and covariance regularization and matches state-of-the-art downstream performance.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"VICReg prevents collapse to constant embeddings in self-supervised learning by adding an explicit variance term per dimension plus covariance regularization.","strongest_claim":"VICReg achieves results on par with the state of the art on several downstream tasks. In addition, we show that incorporating our new variance term into other methods helps stabilize the training and leads to performance improvements.","weakest_assumption":"The assumption that enforcing per-dimension variance above a fixed threshold (combined with covariance regularization) is sufficient to eliminate collapse across architectures and datasets without introducing new failure modes or requiring architecture-specific adjustments."}},"verdict_id":"59a6ffca-94b0-439a-be4e-b6a9f49cc80a"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:5d735262715b61a0d1fc368293de486e01122c4c7bf470e2c67c7927762f8738","target":"record","created_at":"2026-05-17T23:38:51Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"369d5bbba7f095e3dc758d10cbbf40559d032349b243141452acb338d3ae81e9","cross_cats_sorted":["cs.AI","cs.LG"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2021-05-11T09:53:21Z","title_canon_sha256":"8bc0a69c90c72dd6963df5adfd4d9898ef7492bc826dadfa623cd2f91e796c9a"},"schema_version":"1.0","source":{"id":"2105.04906","kind":"arxiv","version":3}},"canonical_sha256":"cee19ad11ea4d198f59d3e21248d685611db12ed366cbcf1f061fe9bdcde2786","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"cee19ad11ea4d198f59d3e21248d685611db12ed366cbcf1f061fe9bdcde2786","first_computed_at":"2026-05-17T23:38:51.093051Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:51.093051Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"+yZq5KRN4VoWk0UkqRRGThxSF4dM8vxyPrhQPU8akyfmfaxML1zTZww38ebx40VW7dml97LYKbDh+Qgq1CTTBQ==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:51.093498Z","signed_message":"canonical_sha256_bytes"},"source_id":"2105.04906","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:5d735262715b61a0d1fc368293de486e01122c4c7bf470e2c67c7927762f8738","sha256:aeb997beb8c2e25fcfde825cdd6fcf0fae2cbf682126c20ce4847396aa6ac9b7"],"state_sha256":"d4a87acae07bf4630ab0acf8fdd6b0b91504534e5c444f58b94b8af86de614c8"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"Ez9DaREMT/KfHgt4pN1ckEtGvgF2OCxMAFh44JAM9vAinJhksvKvK1PfugLjj4GgfqZnCmapwxPoA+Yb5fpYAQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-31T16:28:35.040670Z","bundle_sha256":"c1b192f26818c7d193ed3de0bfd1dafb46ad46e8fb203f0a8340a4e1b6bb149a"}}