{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:KP2GUVDWWQUOFCXGEP3S5E24AH","short_pith_number":"pith:KP2GUVDW","schema_version":"1.0","canonical_sha256":"53f46a5476b428e28ae623f72e935c01fd297950437a8d844a1a384244cf0348","source":{"kind":"arxiv","id":"2310.17813","version":2},"attestation_state":"computed","paper":{"title":"A Spectral Condition for Feature Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Greg Yang, James B. Simon, Jeremy Bernstein","submitted_at":"2023-10-26T23:17:39Z","abstract_excerpt":"The push to train ever larger neural networks has motivated the study of initialization and training at large network width. A key challenge is to scale training so that a network's internal representations evolve nontrivially at all widths, a process known as feature learning. Here, we show that feature learning is achieved by scaling the spectral norm of weight matrices and their updates like $\\sqrt{\\texttt{fan-out}/\\texttt{fan-in}}$, in contrast to widely used but heuristic scalings based on Frobenius norm and entry size. Our spectral scaling analysis also leads to an elementary derivation "},"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":"2310.17813","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2023-10-26T23:17:39Z","cross_cats_sorted":[],"title_canon_sha256":"a1e2fccc127196453e51023270da6f5e7b81983e874bea1606221b3ebc979ab6","abstract_canon_sha256":"37f9933611fe038a23c3b30e8b7a6765fc6bab71b236e3adec6a01d02da3d5ac"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:18:45.527172Z","signature_b64":"ccO3XtV7ki9JDpGeb41Q7JcYb7Q+E+zoJMmdQnkKqm4buZSIB878l7SkXET7EyFCMS1AMDxyTlG6mmBRWOTzAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"53f46a5476b428e28ae623f72e935c01fd297950437a8d844a1a384244cf0348","last_reissued_at":"2026-07-05T08:18:45.526680Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:18:45.526680Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Spectral Condition for Feature Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Greg Yang, James B. Simon, Jeremy Bernstein","submitted_at":"2023-10-26T23:17:39Z","abstract_excerpt":"The push to train ever larger neural networks has motivated the study of initialization and training at large network width. A key challenge is to scale training so that a network's internal representations evolve nontrivially at all widths, a process known as feature learning. Here, we show that feature learning is achieved by scaling the spectral norm of weight matrices and their updates like $\\sqrt{\\texttt{fan-out}/\\texttt{fan-in}}$, in contrast to widely used but heuristic scalings based on Frobenius norm and entry size. Our spectral scaling analysis also leads to an elementary derivation "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2310.17813","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/2310.17813/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":"2310.17813","created_at":"2026-07-05T08:18:45.526744+00:00"},{"alias_kind":"arxiv_version","alias_value":"2310.17813v2","created_at":"2026-07-05T08:18:45.526744+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2310.17813","created_at":"2026-07-05T08:18:45.526744+00:00"},{"alias_kind":"pith_short_12","alias_value":"KP2GUVDWWQUO","created_at":"2026-07-05T08:18:45.526744+00:00"},{"alias_kind":"pith_short_16","alias_value":"KP2GUVDWWQUOFCXG","created_at":"2026-07-05T08:18:45.526744+00:00"},{"alias_kind":"pith_short_8","alias_value":"KP2GUVDW","created_at":"2026-07-05T08:18:45.526744+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":24,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09658","citing_title":"Muon Learns More Robust and Transferable Features than Adam","ref_index":134,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06470","citing_title":"PC Layer: Polynomial Weight Preconditioning for Improving LLM Pre-Training","ref_index":109,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06418","citing_title":"Double Preconditioning (DoPr): Optimization for Test-Time Performance, not Validation Loss","ref_index":107,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05484","citing_title":"Learned Subspace Compression for Communication-Efficient Pipeline Parallelism","ref_index":62,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04662","citing_title":"Why Muon Outperforms Adam: A Curvature Perspective","ref_index":56,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04048","citing_title":"Unlocking Feature Learning in Gated Delta Networks at Scale","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18528","citing_title":"Scale-Invariant Neural Network Optimization: Norm Geometry and Heavy-Tailed Noise","ref_index":133,"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":33,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26459","citing_title":"MuCon: Clipped Muon Updates for LLM Training","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21803","citing_title":"Same Architecture, Different Capacity: Optimizer-Induced Spectral Scaling Laws","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2502.07529","citing_title":"Training Deep Learning Models with Norm-Constrained LMOs","ref_index":214,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18528","citing_title":"Scale-Invariant Neural Network Optimization: Norm Geometry and Heavy-Tailed Noise","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15290","citing_title":"GQA-{\\mu}P: The maximal parameterization update for grouped query attention","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2510.04212","citing_title":"Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2403.03507","citing_title":"GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection","ref_index":50,"is_internal_anchor":false},{"citing_arxiv_id":"2409.20325","citing_title":"Old Optimizer, New Norm: An Anthology","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2603.00541","citing_title":"Spectral Condition for $\\mu$P under Width-Depth Scaling","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02505","citing_title":"Optimal Projection-Free Adaptive SGD for Matrix Optimization","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11872","citing_title":"LOFT: Low-Rank Orthogonal Fine-Tuning via Task-Aware Support Selection","ref_index":42,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12492","citing_title":"Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation","ref_index":86,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09238","citing_title":"Intrinsic Muon: Spectral Optimization on Riemannian Matrix Manifolds","ref_index":68,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10164","citing_title":"Hyperparameter Transfer for Dense Associative Memories","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06654","citing_title":"Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2605.04418","citing_title":"Demystifying Manifold Constraints in LLM Pre-training","ref_index":21,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/KP2GUVDWWQUOFCXGEP3S5E24AH","json":"https://pith.science/pith/KP2GUVDWWQUOFCXGEP3S5E24AH.json","graph_json":"https://pith.science/api/pith-number/KP2GUVDWWQUOFCXGEP3S5E24AH/graph.json","events_json":"https://pith.science/api/pith-number/KP2GUVDWWQUOFCXGEP3S5E24AH/events.json","paper":"https://pith.science/paper/KP2GUVDW"},"agent_actions":{"view_html":"https://pith.science/pith/KP2GUVDWWQUOFCXGEP3S5E24AH","download_json":"https://pith.science/pith/KP2GUVDWWQUOFCXGEP3S5E24AH.json","view_paper":"https://pith.science/paper/KP2GUVDW","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2310.17813&json=true","fetch_graph":"https://pith.science/api/pith-number/KP2GUVDWWQUOFCXGEP3S5E24AH/graph.json","fetch_events":"https://pith.science/api/pith-number/KP2GUVDWWQUOFCXGEP3S5E24AH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/KP2GUVDWWQUOFCXGEP3S5E24AH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/KP2GUVDWWQUOFCXGEP3S5E24AH/action/storage_attestation","attest_author":"https://pith.science/pith/KP2GUVDWWQUOFCXGEP3S5E24AH/action/author_attestation","sign_citation":"https://pith.science/pith/KP2GUVDWWQUOFCXGEP3S5E24AH/action/citation_signature","submit_replication":"https://pith.science/pith/KP2GUVDWWQUOFCXGEP3S5E24AH/action/replication_record"}},"created_at":"2026-07-05T08:18:45.526744+00:00","updated_at":"2026-07-05T08:18:45.526744+00:00"}