{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2019:K62EWTFOSFRZAQTOBXEMZD77SO","short_pith_number":"pith:K62EWTFO","schema_version":"1.0","canonical_sha256":"57b44b4cae916390426e0dc8cc8fff938b3cb59ac348381dab61042be8fbb8ac","source":{"kind":"arxiv","id":"1903.05662","version":4},"attestation_state":"computed","paper":{"title":"Understanding Straight-Through Estimator in Training Activation Quantized Neural Nets","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Jack Xin, Jiancheng Lyu, Penghang Yin, Shuai Zhang, Stanley Osher, Yingyong Qi","submitted_at":"2019-03-13T18:23:43Z","abstract_excerpt":"Training activation quantized neural networks involves minimizing a piecewise constant function whose gradient vanishes almost everywhere, which is undesirable for the standard back-propagation or chain rule. An empirical way around this issue is to use a straight-through estimator (STE) (Bengio et al., 2013) in the backward pass only, so that the \"gradient\" through the modified chain rule becomes non-trivial. Since this unusual \"gradient\" is certainly not the gradient of loss function, the following question arises: why searching in its negative direction minimizes the training loss? In 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":"1903.05662","kind":"arxiv","version":4},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2019-03-13T18:23:43Z","cross_cats_sorted":["math.OC","stat.ML"],"title_canon_sha256":"361294f6ffcba2be605f1106ca29f54a752f600f4a720136fb61cfbc94b89556","abstract_canon_sha256":"d78491da3c8da97fd4e8d13cadd1bfadc9faefc9bc84e55210f0db6f40fdfa2d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T00:07:16.911964Z","signature_b64":"5r43b/JC9uEU4GODYWTif/pJT2iRIcYkNdDaeynJU0nxvRxDFMGUL3rvO467R1rTGAy12mYiYoQAtcvvSPCOAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"57b44b4cae916390426e0dc8cc8fff938b3cb59ac348381dab61042be8fbb8ac","last_reissued_at":"2026-07-05T00:07:16.911455Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T00:07:16.911455Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding Straight-Through Estimator in Training Activation Quantized Neural Nets","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Jack Xin, Jiancheng Lyu, Penghang Yin, Shuai Zhang, Stanley Osher, Yingyong Qi","submitted_at":"2019-03-13T18:23:43Z","abstract_excerpt":"Training activation quantized neural networks involves minimizing a piecewise constant function whose gradient vanishes almost everywhere, which is undesirable for the standard back-propagation or chain rule. An empirical way around this issue is to use a straight-through estimator (STE) (Bengio et al., 2013) in the backward pass only, so that the \"gradient\" through the modified chain rule becomes non-trivial. Since this unusual \"gradient\" is certainly not the gradient of loss function, the following question arises: why searching in its negative direction minimizes the training loss? In this "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"1903.05662","kind":"arxiv","version":4},"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/1903.05662/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":"1903.05662","created_at":"2026-07-05T00:07:16.911513+00:00"},{"alias_kind":"arxiv_version","alias_value":"1903.05662v4","created_at":"2026-07-05T00:07:16.911513+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.1903.05662","created_at":"2026-07-05T00:07:16.911513+00:00"},{"alias_kind":"pith_short_12","alias_value":"K62EWTFOSFRZ","created_at":"2026-07-05T00:07:16.911513+00:00"},{"alias_kind":"pith_short_16","alias_value":"K62EWTFOSFRZAQTO","created_at":"2026-07-05T00:07:16.911513+00:00"},{"alias_kind":"pith_short_8","alias_value":"K62EWTFO","created_at":"2026-07-05T00:07:16.911513+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":14,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25087","citing_title":"Neural Network Quantization by Learning Low-Loss Subspaces","ref_index":54,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10531","citing_title":"LC-QAT: Data-Efficient 2-Bit QAT for LLMs via Linear-Constrained Vector Quantization","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10531","citing_title":"LC-QAT: Data-Efficient 2-Bit QAT for LLMs via Linear-Constrained Vector Quantization","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01127","citing_title":"$\\text{Log}_\\text{b}$Quant: Quantizing Language Models in Logarithmic Space","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15216","citing_title":"Hardware-Software Co-Design of Scalable, Energy-Efficient Analog Recurrent Computations","ref_index":86,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00078","citing_title":"Flow-Based Generative Modeling for Optimizing Sampling Policies in Compressed Sensing Applications","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15216","citing_title":"Hardware-Software Co-Design of Scalable, Energy-Efficient Analog Recurrent Computations","ref_index":86,"is_internal_anchor":false},{"citing_arxiv_id":"2602.15451","citing_title":"Molecular Design beyond Training Data with Novel Extended Objective Functionals of Generative AI Models Driven by Quantum Annealing Computer","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2604.03044","citing_title":"JoyAI-LLM Flash: Advancing Mid-Scale LLMs with Token Efficiency","ref_index":86,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11558","citing_title":"A Composite Activation Function for Learning Stable Binary Representations","ref_index":77,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09416","citing_title":"A Controlled Diagnostic Study of Hardware-Induced Distortions in Hardware-Aware Training","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09256","citing_title":"Improving Generalization by Permutation Routing Across Model Copies","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19167","citing_title":"LBLLM: Lightweight Binarization of Large Language Models via Three-Stage Distillation","ref_index":89,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10861","citing_title":"Training single-electron and single-photon stochastic physical neural networks","ref_index":36,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/K62EWTFOSFRZAQTOBXEMZD77SO","json":"https://pith.science/pith/K62EWTFOSFRZAQTOBXEMZD77SO.json","graph_json":"https://pith.science/api/pith-number/K62EWTFOSFRZAQTOBXEMZD77SO/graph.json","events_json":"https://pith.science/api/pith-number/K62EWTFOSFRZAQTOBXEMZD77SO/events.json","paper":"https://pith.science/paper/K62EWTFO"},"agent_actions":{"view_html":"https://pith.science/pith/K62EWTFOSFRZAQTOBXEMZD77SO","download_json":"https://pith.science/pith/K62EWTFOSFRZAQTOBXEMZD77SO.json","view_paper":"https://pith.science/paper/K62EWTFO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=1903.05662&json=true","fetch_graph":"https://pith.science/api/pith-number/K62EWTFOSFRZAQTOBXEMZD77SO/graph.json","fetch_events":"https://pith.science/api/pith-number/K62EWTFOSFRZAQTOBXEMZD77SO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/K62EWTFOSFRZAQTOBXEMZD77SO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/K62EWTFOSFRZAQTOBXEMZD77SO/action/storage_attestation","attest_author":"https://pith.science/pith/K62EWTFOSFRZAQTOBXEMZD77SO/action/author_attestation","sign_citation":"https://pith.science/pith/K62EWTFOSFRZAQTOBXEMZD77SO/action/citation_signature","submit_replication":"https://pith.science/pith/K62EWTFOSFRZAQTOBXEMZD77SO/action/replication_record"}},"created_at":"2026-07-05T00:07:16.911513+00:00","updated_at":"2026-07-05T00:07:16.911513+00:00"}