{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:JELNO6UHUCFK3R4OPIAPXEWL42","short_pith_number":"pith:JELNO6UH","schema_version":"1.0","canonical_sha256":"4916d77a87a08aadc78e7a00fb92cbe6b13c8fca1996c9d9214a356c1c117c67","source":{"kind":"arxiv","id":"2411.14517","version":3},"attestation_state":"computed","paper":{"title":"The Double-Ellipsoid Geometry of CLIP","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Guy Gilboa, Meir Yossef Levi","submitted_at":"2024-11-21T16:27:22Z","abstract_excerpt":"Contrastive Language-Image Pre-Training (CLIP) is highly instrumental in machine learning applications within a large variety of domains. We investigate the geometry of this embedding, which is still not well understood. We examine the raw unnormalized embedding and show that text and image reside on linearly separable ellipsoid shells, not centered at the origin. We explain the benefits of having this structure, allowing to better embed instances according to their uncertainty during contrastive training. Frequent concepts in the dataset yield more false negatives, inducing greater uncertaint"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2411.14517","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2024-11-21T16:27:22Z","cross_cats_sorted":[],"title_canon_sha256":"04f2d5c8d5c3fd4170b4ee37d04b71148fdb106a077c84c8845edba819d42f7c","abstract_canon_sha256":"0a05808487d8333de5c7e29dccb72fbf0657ca2e1bd93283ebd74bd6330dd1cb"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:09:04.330296Z","signature_b64":"HO3QVPrLhfDM4ld33w/8i4kHv8G3hQy6OIqqoAvwCi/7CdyLWpPwxx5tI7WJ65tzmyfFAv/yp0MkOCnuOlBsBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4916d77a87a08aadc78e7a00fb92cbe6b13c8fca1996c9d9214a356c1c117c67","last_reissued_at":"2026-07-05T11:09:04.329884Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:09:04.329884Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Double-Ellipsoid Geometry of CLIP","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Guy Gilboa, Meir Yossef Levi","submitted_at":"2024-11-21T16:27:22Z","abstract_excerpt":"Contrastive Language-Image Pre-Training (CLIP) is highly instrumental in machine learning applications within a large variety of domains. We investigate the geometry of this embedding, which is still not well understood. We examine the raw unnormalized embedding and show that text and image reside on linearly separable ellipsoid shells, not centered at the origin. We explain the benefits of having this structure, allowing to better embed instances according to their uncertainty during contrastive training. Frequent concepts in the dataset yield more false negatives, inducing greater uncertaint"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2411.14517","kind":"arxiv","version":3},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2411.14517/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2411.14517","created_at":"2026-07-05T11:09:04.329940+00:00"},{"alias_kind":"arxiv_version","alias_value":"2411.14517v3","created_at":"2026-07-05T11:09:04.329940+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2411.14517","created_at":"2026-07-05T11:09:04.329940+00:00"},{"alias_kind":"pith_short_12","alias_value":"JELNO6UHUCFK","created_at":"2026-07-05T11:09:04.329940+00:00"},{"alias_kind":"pith_short_16","alias_value":"JELNO6UHUCFK3R4O","created_at":"2026-07-05T11:09:04.329940+00:00"},{"alias_kind":"pith_short_8","alias_value":"JELNO6UH","created_at":"2026-07-05T11:09:04.329940+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.30625","citing_title":"Optimization Dynamics Imprint Semantic Specificity in Contrastive Embedding Norms","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2606.11190","citing_title":"When to Align, When to Predict: A Phase Diagram for Multimodal Learning","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07907","citing_title":"Consistency Regularised Gradient Flows for Inverse Problems","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07973","citing_title":"HEART: Hyperspherical Embedding Alignment via Kent-Representation Traversal in Diffusion Models","ref_index":5,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/JELNO6UHUCFK3R4OPIAPXEWL42","json":"https://pith.science/pith/JELNO6UHUCFK3R4OPIAPXEWL42.json","graph_json":"https://pith.science/api/pith-number/JELNO6UHUCFK3R4OPIAPXEWL42/graph.json","events_json":"https://pith.science/api/pith-number/JELNO6UHUCFK3R4OPIAPXEWL42/events.json","paper":"https://pith.science/paper/JELNO6UH"},"agent_actions":{"view_html":"https://pith.science/pith/JELNO6UHUCFK3R4OPIAPXEWL42","download_json":"https://pith.science/pith/JELNO6UHUCFK3R4OPIAPXEWL42.json","view_paper":"https://pith.science/paper/JELNO6UH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2411.14517&json=true","fetch_graph":"https://pith.science/api/pith-number/JELNO6UHUCFK3R4OPIAPXEWL42/graph.json","fetch_events":"https://pith.science/api/pith-number/JELNO6UHUCFK3R4OPIAPXEWL42/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/JELNO6UHUCFK3R4OPIAPXEWL42/action/timestamp_anchor","attest_storage":"https://pith.science/pith/JELNO6UHUCFK3R4OPIAPXEWL42/action/storage_attestation","attest_author":"https://pith.science/pith/JELNO6UHUCFK3R4OPIAPXEWL42/action/author_attestation","sign_citation":"https://pith.science/pith/JELNO6UHUCFK3R4OPIAPXEWL42/action/citation_signature","submit_replication":"https://pith.science/pith/JELNO6UHUCFK3R4OPIAPXEWL42/action/replication_record"}},"created_at":"2026-07-05T11:09:04.329940+00:00","updated_at":"2026-07-05T11:09:04.329940+00:00"}