{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2018:2FGXYWYDWCT3XWECO4BFZ5ZK6E","short_pith_number":"pith:2FGXYWYD","canonical_record":{"source":{"id":"1801.01401","kind":"arxiv","version":5},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"stat.ML","submitted_at":"2018-01-04T15:25:26Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"fd2357f919bc86340dbd2aa08f818090f09a7efd61582570a969498869ebeaf0","abstract_canon_sha256":"d2c6c23e9048f2e7f953dc43071d15ee896a0710d38fa54558867dd0f7438474"},"schema_version":"1.0"},"canonical_sha256":"d14d7c5b03b0a7bbd88277025cf72af1169c14156463487d5369739d8cc42e61","source":{"kind":"arxiv","id":"1801.01401","version":5},"source_aliases":[{"alias_kind":"arxiv","alias_value":"1801.01401","created_at":"2026-05-17T23:39:05Z"},{"alias_kind":"arxiv_version","alias_value":"1801.01401v5","created_at":"2026-05-17T23:39:05Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.1801.01401","created_at":"2026-05-17T23:39:05Z"},{"alias_kind":"pith_short_12","alias_value":"2FGXYWYDWCT3","created_at":"2026-05-18T12:32:02Z"},{"alias_kind":"pith_short_16","alias_value":"2FGXYWYDWCT3XWEC","created_at":"2026-05-18T12:32:02Z"},{"alias_kind":"pith_short_8","alias_value":"2FGXYWYD","created_at":"2026-05-18T12:32:02Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2018:2FGXYWYDWCT3XWECO4BFZ5ZK6E","target":"record","payload":{"canonical_record":{"source":{"id":"1801.01401","kind":"arxiv","version":5},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"stat.ML","submitted_at":"2018-01-04T15:25:26Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"fd2357f919bc86340dbd2aa08f818090f09a7efd61582570a969498869ebeaf0","abstract_canon_sha256":"d2c6c23e9048f2e7f953dc43071d15ee896a0710d38fa54558867dd0f7438474"},"schema_version":"1.0"},"canonical_sha256":"d14d7c5b03b0a7bbd88277025cf72af1169c14156463487d5369739d8cc42e61","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:05.169414Z","signature_b64":"ZrED+Q8wzcx6pUtYQ/CTyNNT2oqtPlJ0fZKxfHjPTGFRA0UtVMYD7zHylsrzg+XYVA249VhZTMq9fKBrrvgxBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d14d7c5b03b0a7bbd88277025cf72af1169c14156463487d5369739d8cc42e61","last_reissued_at":"2026-05-17T23:39:05.168704Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:05.168704Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"1801.01401","source_version":5,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:39:05Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"JtwO6NmDInlVEDOjz+okcmQsDuplBsXPY7xWGRy4WxlJQ/P+JB7evN63IZcuRjIINKHBU/8rpbCorbkZgWfiAg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-04T08:05:16.169850Z"},"content_sha256":"acba0b1dac62fb2449fcaa80001ad8fddde1c0206e83d1ae42a7cdb4a745be27","schema_version":"1.0","event_id":"sha256:acba0b1dac62fb2449fcaa80001ad8fddde1c0206e83d1ae42a7cdb4a745be27"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2018:2FGXYWYDWCT3XWECO4BFZ5ZK6E","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Demystifying MMD GANs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Gradient estimators for MMD GANs and Wasserstein GANs are unbiased, but finite-sample discriminators bias the generator updates.","cross_cats":["cs.LG"],"primary_cat":"stat.ML","authors_text":"Arthur Gretton, Danica J. Sutherland, Michael Arbel, Miko{\\l}aj Bi\\'nkowski","submitted_at":"2018-01-04T15:25:26Z","abstract_excerpt":"We investigate the training and performance of generative adversarial networks using the Maximum Mean Discrepancy (MMD) as critic, termed MMD GANs. As our main theoretical contribution, we clarify the situation with bias in GAN loss functions raised by recent work: we show that gradient estimators used in the optimization process for both MMD GANs and Wasserstein GANs are unbiased, but learning a discriminator based on samples leads to biased gradients for the generator parameters. We also discuss the issue of kernel choice for the MMD critic, and characterize the kernel corresponding to the e"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We show that gradient estimators used in the optimization process for both MMD GANs and Wasserstein GANs are unbiased, but learning a discriminator based on samples leads to biased gradients for the generator parameters.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The theoretical unbiasedness of the critic gradients holds under the assumption that the kernel is fixed and positive definite, and that the practical bias from finite samples does not dominate other optimization issues in real training.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MMD GANs have unbiased critic gradients but biased generator gradients from sample-based learning, and the Kernel Inception Distance provides a practical new measure for GAN convergence and dynamic learning rate adaptation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Gradient estimators for MMD GANs and Wasserstein GANs are unbiased, but finite-sample discriminators bias the generator updates.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"81369845aaad7266f9e364ca18e79093c6788bbb11d976b27749d1fa750f63af"},"source":{"id":"1801.01401","kind":"arxiv","version":5},"verdict":{"id":"83dac35c-c866-4257-b03f-e46451bfcb13","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:03:14.672850Z","strongest_claim":"We show that gradient estimators used in the optimization process for both MMD GANs and Wasserstein GANs are unbiased, but learning a discriminator based on samples leads to biased gradients for the generator parameters.","one_line_summary":"MMD GANs have unbiased critic gradients but biased generator gradients from sample-based learning, and the Kernel Inception Distance provides a practical new measure for GAN convergence and dynamic learning rate adaptation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The theoretical unbiasedness of the critic gradients holds under the assumption that the kernel is fixed and positive definite, and that the practical bias from finite samples does not dominate other optimization issues in real training.","pith_extraction_headline":"Gradient estimators for MMD GANs and Wasserstein GANs are unbiased, but finite-sample discriminators bias the generator updates."},"references":{"count":62,"sample":[{"doi":"","year":2017,"title":"arXiv preprint arXiv:1701.04862 , year=","work_id":"965f2446-b480-4da4-81ba-3df9c5ea4efc","ref_index":1,"cited_arxiv_id":"1701.04862","is_internal_anchor":true},{"doi":"","year":2017,"title":"Wasserstein GAN","work_id":"328412c6-1a94-4a20-b3f4-af6a7986efee","ref_index":2,"cited_arxiv_id":"1701.07875","is_internal_anchor":true},{"doi":"","year":2017,"title":"Do GAN s actually learn the distribution? an empirical study","work_id":"58db00de-2a73-41b1-aa81-e337a4745825","ref_index":3,"cited_arxiv_id":"1706.08224","is_internal_anchor":true},{"doi":"","year":2017,"title":"Generalization and Equilibrium in Generative Adversarial Nets (GANs)","work_id":"a88096c3-20f4-4208-8293-fa40bef46448","ref_index":4,"cited_arxiv_id":"1703.00573","is_internal_anchor":true},{"doi":"","year":2017,"title":"Bellemare, Ivo Danihelka, Will Dabney, Shakir Mohamed, Balaji Lakshminarayanan, Stephan Hoyer, and R´emi Munos","work_id":"5dfdb668-ff19-4c96-acc8-bc352a3ec469","ref_index":5,"cited_arxiv_id":"1705.10743","is_internal_anchor":true}],"resolved_work":62,"snapshot_sha256":"60b46ab3c3fc33be876ca974192541e7397af4f55ab7116491dd105567d01f0f","internal_anchors":39},"formal_canon":{"evidence_count":1,"snapshot_sha256":"056359fe4801bf883afdb631c971cd94b393fcf0fb16ab0d5295aa49a8c1f1f8"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"83dac35c-c866-4257-b03f-e46451bfcb13"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:39:05Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"AZwZkLU2Sijbm59L5lt2ouKyrrIaum5KGml/h9SbL2qdvMDycJugyM7OvkAwz89jMUdO4PqdiykPtAEEjrh/Bw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-04T08:05:16.170421Z"},"content_sha256":"46adb35eee7968c8c0f52a5447799f8a5e6ddb103bb5c8fed6d1ae9a65c87d96","schema_version":"1.0","event_id":"sha256:46adb35eee7968c8c0f52a5447799f8a5e6ddb103bb5c8fed6d1ae9a65c87d96"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/2FGXYWYDWCT3XWECO4BFZ5ZK6E/bundle.json","state_url":"https://pith.science/pith/2FGXYWYDWCT3XWECO4BFZ5ZK6E/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/2FGXYWYDWCT3XWECO4BFZ5ZK6E/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":"2026-08-04T08:05:16Z","links":{"resolver":"https://pith.science/pith/2FGXYWYDWCT3XWECO4BFZ5ZK6E","bundle":"https://pith.science/pith/2FGXYWYDWCT3XWECO4BFZ5ZK6E/bundle.json","state":"https://pith.science/pith/2FGXYWYDWCT3XWECO4BFZ5ZK6E/state.json","well_known_bundle":"https://pith.science/.well-known/pith/2FGXYWYDWCT3XWECO4BFZ5ZK6E/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2018:2FGXYWYDWCT3XWECO4BFZ5ZK6E","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":"d2c6c23e9048f2e7f953dc43071d15ee896a0710d38fa54558867dd0f7438474","cross_cats_sorted":["cs.LG"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"stat.ML","submitted_at":"2018-01-04T15:25:26Z","title_canon_sha256":"fd2357f919bc86340dbd2aa08f818090f09a7efd61582570a969498869ebeaf0"},"schema_version":"1.0","source":{"id":"1801.01401","kind":"arxiv","version":5}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"1801.01401","created_at":"2026-05-17T23:39:05Z"},{"alias_kind":"arxiv_version","alias_value":"1801.01401v5","created_at":"2026-05-17T23:39:05Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.1801.01401","created_at":"2026-05-17T23:39:05Z"},{"alias_kind":"pith_short_12","alias_value":"2FGXYWYDWCT3","created_at":"2026-05-18T12:32:02Z"},{"alias_kind":"pith_short_16","alias_value":"2FGXYWYDWCT3XWEC","created_at":"2026-05-18T12:32:02Z"},{"alias_kind":"pith_short_8","alias_value":"2FGXYWYD","created_at":"2026-05-18T12:32:02Z"}],"graph_snapshots":[{"event_id":"sha256:46adb35eee7968c8c0f52a5447799f8a5e6ddb103bb5c8fed6d1ae9a65c87d96","target":"graph","created_at":"2026-05-17T23:39:05Z","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":"We show that gradient estimators used in the optimization process for both MMD GANs and Wasserstein GANs are unbiased, but learning a discriminator based on samples leads to biased gradients for the generator parameters."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"The theoretical unbiasedness of the critic gradients holds under the assumption that the kernel is fixed and positive definite, and that the practical bias from finite samples does not dominate other optimization issues in real training."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"MMD GANs have unbiased critic gradients but biased generator gradients from sample-based learning, and the Kernel Inception Distance provides a practical new measure for GAN convergence and dynamic learning rate adaptation."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Gradient estimators for MMD GANs and Wasserstein GANs are unbiased, but finite-sample discriminators bias the generator updates."}],"snapshot_sha256":"81369845aaad7266f9e364ca18e79093c6788bbb11d976b27749d1fa750f63af"},"formal_canon":{"evidence_count":1,"snapshot_sha256":"056359fe4801bf883afdb631c971cd94b393fcf0fb16ab0d5295aa49a8c1f1f8"},"paper":{"abstract_excerpt":"We investigate the training and performance of generative adversarial networks using the Maximum Mean Discrepancy (MMD) as critic, termed MMD GANs. As our main theoretical contribution, we clarify the situation with bias in GAN loss functions raised by recent work: we show that gradient estimators used in the optimization process for both MMD GANs and Wasserstein GANs are unbiased, but learning a discriminator based on samples leads to biased gradients for the generator parameters. We also discuss the issue of kernel choice for the MMD critic, and characterize the kernel corresponding to the e","authors_text":"Arthur Gretton, Danica J. Sutherland, Michael Arbel, Miko{\\l}aj Bi\\'nkowski","cross_cats":["cs.LG"],"headline":"Gradient estimators for MMD GANs and Wasserstein GANs are unbiased, but finite-sample discriminators bias the generator updates.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"stat.ML","submitted_at":"2018-01-04T15:25:26Z","title":"Demystifying MMD GANs"},"references":{"count":62,"internal_anchors":39,"resolved_work":62,"sample":[{"cited_arxiv_id":"1701.04862","doi":"","is_internal_anchor":true,"ref_index":1,"title":"arXiv preprint arXiv:1701.04862 , year=","work_id":"965f2446-b480-4da4-81ba-3df9c5ea4efc","year":2017},{"cited_arxiv_id":"1701.07875","doi":"","is_internal_anchor":true,"ref_index":2,"title":"Wasserstein GAN","work_id":"328412c6-1a94-4a20-b3f4-af6a7986efee","year":2017},{"cited_arxiv_id":"1706.08224","doi":"","is_internal_anchor":true,"ref_index":3,"title":"Do GAN s actually learn the distribution? an empirical study","work_id":"58db00de-2a73-41b1-aa81-e337a4745825","year":2017},{"cited_arxiv_id":"1703.00573","doi":"","is_internal_anchor":true,"ref_index":4,"title":"Generalization and Equilibrium in Generative Adversarial Nets (GANs)","work_id":"a88096c3-20f4-4208-8293-fa40bef46448","year":2017},{"cited_arxiv_id":"1705.10743","doi":"","is_internal_anchor":true,"ref_index":5,"title":"Bellemare, Ivo Danihelka, Will Dabney, Shakir Mohamed, Balaji Lakshminarayanan, Stephan Hoyer, and R´emi Munos","work_id":"5dfdb668-ff19-4c96-acc8-bc352a3ec469","year":2017}],"snapshot_sha256":"60b46ab3c3fc33be876ca974192541e7397af4f55ab7116491dd105567d01f0f"},"source":{"id":"1801.01401","kind":"arxiv","version":5},"verdict":{"created_at":"2026-05-15T01:03:14.672850Z","id":"83dac35c-c866-4257-b03f-e46451bfcb13","model_set":{"reader":"grok-4.3"},"one_line_summary":"MMD GANs have unbiased critic gradients but biased generator gradients from sample-based learning, and the Kernel Inception Distance provides a practical new measure for GAN convergence and dynamic learning rate adaptation.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Gradient estimators for MMD GANs and Wasserstein GANs are unbiased, but finite-sample discriminators bias the generator updates.","strongest_claim":"We show that gradient estimators used in the optimization process for both MMD GANs and Wasserstein GANs are unbiased, but learning a discriminator based on samples leads to biased gradients for the generator parameters.","weakest_assumption":"The theoretical unbiasedness of the critic gradients holds under the assumption that the kernel is fixed and positive definite, and that the practical bias from finite samples does not dominate other optimization issues in real training."}},"verdict_id":"83dac35c-c866-4257-b03f-e46451bfcb13"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:acba0b1dac62fb2449fcaa80001ad8fddde1c0206e83d1ae42a7cdb4a745be27","target":"record","created_at":"2026-05-17T23:39:05Z","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":"d2c6c23e9048f2e7f953dc43071d15ee896a0710d38fa54558867dd0f7438474","cross_cats_sorted":["cs.LG"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"stat.ML","submitted_at":"2018-01-04T15:25:26Z","title_canon_sha256":"fd2357f919bc86340dbd2aa08f818090f09a7efd61582570a969498869ebeaf0"},"schema_version":"1.0","source":{"id":"1801.01401","kind":"arxiv","version":5}},"canonical_sha256":"d14d7c5b03b0a7bbd88277025cf72af1169c14156463487d5369739d8cc42e61","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"d14d7c5b03b0a7bbd88277025cf72af1169c14156463487d5369739d8cc42e61","first_computed_at":"2026-05-17T23:39:05.168704Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:39:05.168704Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"ZrED+Q8wzcx6pUtYQ/CTyNNT2oqtPlJ0fZKxfHjPTGFRA0UtVMYD7zHylsrzg+XYVA249VhZTMq9fKBrrvgxBQ==","signature_status":"signed_v1","signed_at":"2026-05-17T23:39:05.169414Z","signed_message":"canonical_sha256_bytes"},"source_id":"1801.01401","source_kind":"arxiv","source_version":5}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:acba0b1dac62fb2449fcaa80001ad8fddde1c0206e83d1ae42a7cdb4a745be27","sha256:46adb35eee7968c8c0f52a5447799f8a5e6ddb103bb5c8fed6d1ae9a65c87d96"],"state_sha256":"590dcf148352d9c470bb80ff9205c15658edd93599458e22f0ad42ae26c81d81"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"HDK1btomyIbhFpQT9LMNiHXpzZDcKtwh6mbkEn++leL47yFKLI6GBpeKKYS/ZXpDKjvSh6kuct0WlaPf5ccqAw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-04T08:05:16.173517Z","bundle_sha256":"f2e791d38f6c1a167adb7ddafceea790e970b695c3650554c50701108dfca29c"}}