{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2018:P2ARL6MFRUQQ7AXEAN6PYUQJBA","short_pith_number":"pith:P2ARL6MF","schema_version":"1.0","canonical_sha256":"7e8115f9858d210f82e4037cfc52090830423244577ef53aab0d7cef007cba66","source":{"kind":"arxiv","id":"1805.12177","version":4},"attestation_state":"computed","paper":{"title":"Why do deep convolutional networks generalize so poorly to small image transformations?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Aharon Azulay, Yair Weiss","submitted_at":"2018-05-30T18:56:33Z","abstract_excerpt":"Convolutional Neural Networks (CNNs) are commonly assumed to be invariant to small image transformations: either because of the convolutional architecture or because they were trained using data augmentation. Recently, several authors have shown that this is not the case: small translations or rescalings of the input image can drastically change the network's prediction. In this paper, we quantify this phenomena and ask why neither the convolutional architecture nor data augmentation are sufficient to achieve the desired invariance. Specifically, we show that the convolutional architecture doe"},"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":"1805.12177","kind":"arxiv","version":4},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2018-05-30T18:56:33Z","cross_cats_sorted":[],"title_canon_sha256":"8a3fedccb6683b2c1414c6dd1fd362bfe54e76535c7cfa3b9ce2e1d2ebd6bddc","abstract_canon_sha256":"79324376869d542c6bcc86f681751776f383b4252951d2a52817c4e453452ae5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T00:35:09.089855Z","signature_b64":"lH6s4kaa46F+fcheL5xqsFmvaUwh2y7oe468q5FDOWrA1tuylxaPh9eqGUQ8Mz9jCNUr5o6FVIRrduHsDoTzCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7e8115f9858d210f82e4037cfc52090830423244577ef53aab0d7cef007cba66","last_reissued_at":"2026-07-05T00:35:09.089432Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T00:35:09.089432Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Why do deep convolutional networks generalize so poorly to small image transformations?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Aharon Azulay, Yair Weiss","submitted_at":"2018-05-30T18:56:33Z","abstract_excerpt":"Convolutional Neural Networks (CNNs) are commonly assumed to be invariant to small image transformations: either because of the convolutional architecture or because they were trained using data augmentation. Recently, several authors have shown that this is not the case: small translations or rescalings of the input image can drastically change the network's prediction. In this paper, we quantify this phenomena and ask why neither the convolutional architecture nor data augmentation are sufficient to achieve the desired invariance. Specifically, we show that the convolutional architecture doe"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"1805.12177","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/1805.12177/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":"1805.12177","created_at":"2026-07-05T00:35:09.089489+00:00"},{"alias_kind":"arxiv_version","alias_value":"1805.12177v4","created_at":"2026-07-05T00:35:09.089489+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.1805.12177","created_at":"2026-07-05T00:35:09.089489+00:00"},{"alias_kind":"pith_short_12","alias_value":"P2ARL6MFRUQQ","created_at":"2026-07-05T00:35:09.089489+00:00"},{"alias_kind":"pith_short_16","alias_value":"P2ARL6MFRUQQ7AXE","created_at":"2026-07-05T00:35:09.089489+00:00"},{"alias_kind":"pith_short_8","alias_value":"P2ARL6MF","created_at":"2026-07-05T00:35:09.089489+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.06624","citing_title":"Principles and Practice of Deep Representation Learning: or a Mathematical Theory of Memory","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27870","citing_title":"Parameter-Efficient Architectural Modifications for Translation-Invariant CNNs","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/P2ARL6MFRUQQ7AXEAN6PYUQJBA","json":"https://pith.science/pith/P2ARL6MFRUQQ7AXEAN6PYUQJBA.json","graph_json":"https://pith.science/api/pith-number/P2ARL6MFRUQQ7AXEAN6PYUQJBA/graph.json","events_json":"https://pith.science/api/pith-number/P2ARL6MFRUQQ7AXEAN6PYUQJBA/events.json","paper":"https://pith.science/paper/P2ARL6MF"},"agent_actions":{"view_html":"https://pith.science/pith/P2ARL6MFRUQQ7AXEAN6PYUQJBA","download_json":"https://pith.science/pith/P2ARL6MFRUQQ7AXEAN6PYUQJBA.json","view_paper":"https://pith.science/paper/P2ARL6MF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=1805.12177&json=true","fetch_graph":"https://pith.science/api/pith-number/P2ARL6MFRUQQ7AXEAN6PYUQJBA/graph.json","fetch_events":"https://pith.science/api/pith-number/P2ARL6MFRUQQ7AXEAN6PYUQJBA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/P2ARL6MFRUQQ7AXEAN6PYUQJBA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/P2ARL6MFRUQQ7AXEAN6PYUQJBA/action/storage_attestation","attest_author":"https://pith.science/pith/P2ARL6MFRUQQ7AXEAN6PYUQJBA/action/author_attestation","sign_citation":"https://pith.science/pith/P2ARL6MFRUQQ7AXEAN6PYUQJBA/action/citation_signature","submit_replication":"https://pith.science/pith/P2ARL6MFRUQQ7AXEAN6PYUQJBA/action/replication_record"}},"created_at":"2026-07-05T00:35:09.089489+00:00","updated_at":"2026-07-05T00:35:09.089489+00:00"}