{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2021:FHXPVUW7Y4LYFIU2FJTUATXQTR","short_pith_number":"pith:FHXPVUW7","schema_version":"1.0","canonical_sha256":"29eefad2dfc71782a29a2a67404ef09c75040e6c8ac595c3ccf586268298c658","source":{"kind":"arxiv","id":"2106.15933","version":2},"attestation_state":"computed","paper":{"title":"Saddle-to-Saddle Dynamics in Deep Linear Networks: Small Initialization Training, Symmetry, and Sparsity","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"stat.ML","authors_text":"Arthur Jacot, Berfin \\c{S}im\\c{s}ek, Cl\\'ement Hongler, Fran\\c{c}ois Ged, Franck Gabriel","submitted_at":"2021-06-30T09:34:05Z","abstract_excerpt":"The dynamics of Deep Linear Networks (DLNs) is dramatically affected by the variance $\\sigma^2$ of the parameters at initialization $\\theta_0$. For DLNs of width $w$, we show a phase transition w.r.t. the scaling $\\gamma$ of the variance $\\sigma^2=w^{-\\gamma}$ as $w\\to\\infty$: for large variance ($\\gamma<1$), $\\theta_0$ is very close to a global minimum but far from any saddle point, and for small variance ($\\gamma>1$), $\\theta_0$ is close to a saddle point and far from any global minimum. While the first case corresponds to the well-studied NTK regime, the second case is less understood. This"},"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":"2106.15933","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"stat.ML","submitted_at":"2021-06-30T09:34:05Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"38930fbf67221358248ad7c6b9badbf4834e81303a1b9d11b63996c210cb83cf","abstract_canon_sha256":"c0488ad553ac2023956c7c4a020d79eff86c503e35ba88d54bb5164e8c731c0d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T03:52:29.620818Z","signature_b64":"rCUXbvAI/khEzZXZRudu/5aX1htJKjPkOQN2ncvurVt3ySIQOgQt9YkwN+Sxjf2OKx+vKe+dQBM0jjX94i6uDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"29eefad2dfc71782a29a2a67404ef09c75040e6c8ac595c3ccf586268298c658","last_reissued_at":"2026-07-05T03:52:29.620354Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T03:52:29.620354Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Saddle-to-Saddle Dynamics in Deep Linear Networks: Small Initialization Training, Symmetry, and Sparsity","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"stat.ML","authors_text":"Arthur Jacot, Berfin \\c{S}im\\c{s}ek, Cl\\'ement Hongler, Fran\\c{c}ois Ged, Franck Gabriel","submitted_at":"2021-06-30T09:34:05Z","abstract_excerpt":"The dynamics of Deep Linear Networks (DLNs) is dramatically affected by the variance $\\sigma^2$ of the parameters at initialization $\\theta_0$. For DLNs of width $w$, we show a phase transition w.r.t. the scaling $\\gamma$ of the variance $\\sigma^2=w^{-\\gamma}$ as $w\\to\\infty$: for large variance ($\\gamma<1$), $\\theta_0$ is very close to a global minimum but far from any saddle point, and for small variance ($\\gamma>1$), $\\theta_0$ is close to a saddle point and far from any global minimum. While the first case corresponds to the well-studied NTK regime, the second case is less understood. This"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2106.15933","kind":"arxiv","version":2},"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/2106.15933/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":"2106.15933","created_at":"2026-07-05T03:52:29.620412+00:00"},{"alias_kind":"arxiv_version","alias_value":"2106.15933v2","created_at":"2026-07-05T03:52:29.620412+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2106.15933","created_at":"2026-07-05T03:52:29.620412+00:00"},{"alias_kind":"pith_short_12","alias_value":"FHXPVUW7Y4LY","created_at":"2026-07-05T03:52:29.620412+00:00"},{"alias_kind":"pith_short_16","alias_value":"FHXPVUW7Y4LYFIU2","created_at":"2026-07-05T03:52:29.620412+00:00"},{"alias_kind":"pith_short_8","alias_value":"FHXPVUW7","created_at":"2026-07-05T03:52:29.620412+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":12,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07884","citing_title":"Optimal Learning Rate Scaling Depends on Data in Deep Scalar Linear Networks","ref_index":33,"is_internal_anchor":true},{"citing_arxiv_id":"2606.23198","citing_title":"Incremental Learning in Mirror Flows","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2607.00665","citing_title":"Effective dynamics of the Sinkhorn algorithm in the regime of low entropy regularization","ref_index":132,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01288","citing_title":"A Theory of Saddle Escape in Deep Nonlinear Networks","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30509","citing_title":"Muon learns balanced solutions in matrix factorization without slow saddle-to-saddle dynamics","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.29548","citing_title":"Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention","ref_index":59,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23087","citing_title":"The Implicit Bias of Depth: From Neural Collapse to Softmax Codes","ref_index":111,"is_internal_anchor":false},{"citing_arxiv_id":"2206.00939","citing_title":"Gradient flow dynamics of shallow ReLU networks for square loss and orthogonal inputs","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01288","citing_title":"A Theory of Saddle Escape in Deep Nonlinear Networks","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2604.21691","citing_title":"There Will Be a Scientific Theory of Deep Learning","ref_index":139,"is_internal_anchor":false},{"citing_arxiv_id":"2604.21905","citing_title":"Low-Rank Adaptation Redux for Large Models","ref_index":76,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01288","citing_title":"A Theory of Saddle Escape in Deep Nonlinear Networks","ref_index":23,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/FHXPVUW7Y4LYFIU2FJTUATXQTR","json":"https://pith.science/pith/FHXPVUW7Y4LYFIU2FJTUATXQTR.json","graph_json":"https://pith.science/api/pith-number/FHXPVUW7Y4LYFIU2FJTUATXQTR/graph.json","events_json":"https://pith.science/api/pith-number/FHXPVUW7Y4LYFIU2FJTUATXQTR/events.json","paper":"https://pith.science/paper/FHXPVUW7"},"agent_actions":{"view_html":"https://pith.science/pith/FHXPVUW7Y4LYFIU2FJTUATXQTR","download_json":"https://pith.science/pith/FHXPVUW7Y4LYFIU2FJTUATXQTR.json","view_paper":"https://pith.science/paper/FHXPVUW7","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2106.15933&json=true","fetch_graph":"https://pith.science/api/pith-number/FHXPVUW7Y4LYFIU2FJTUATXQTR/graph.json","fetch_events":"https://pith.science/api/pith-number/FHXPVUW7Y4LYFIU2FJTUATXQTR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FHXPVUW7Y4LYFIU2FJTUATXQTR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FHXPVUW7Y4LYFIU2FJTUATXQTR/action/storage_attestation","attest_author":"https://pith.science/pith/FHXPVUW7Y4LYFIU2FJTUATXQTR/action/author_attestation","sign_citation":"https://pith.science/pith/FHXPVUW7Y4LYFIU2FJTUATXQTR/action/citation_signature","submit_replication":"https://pith.science/pith/FHXPVUW7Y4LYFIU2FJTUATXQTR/action/replication_record"}},"created_at":"2026-07-05T03:52:29.620412+00:00","updated_at":"2026-07-05T03:52:29.620412+00:00"}