{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:U6J2OD54ACYHPFBJWAECFPES4P","short_pith_number":"pith:U6J2OD54","schema_version":"1.0","canonical_sha256":"a793a70fbc00b0779429b00822bc92e3e181b1c1a3ef2a907fb50e73c0b47cc2","source":{"kind":"arxiv","id":"2605.13835","version":1},"attestation_state":"computed","paper":{"title":"Unlocking Patch-Level Features for CLIP-Based Class-Incremental Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Aligning CLIP patch features to semantic descriptions improves class-incremental learning.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Da-Wei Zhou, Hao Sun, Zi-Jun Ding","submitted_at":"2026-05-13T17:56:23Z","abstract_excerpt":"Class-Incremental Learning (CIL) enables models to continuously integrate new knowledge while mitigating catastrophic forgetting. Driven by the remarkable generalization of CLIP, leveraging pre-trained vision-language models has become a dominant paradigm in CIL. However, current work primarily focuses on aligning global image embeddings (i.e., [CLS] token) with their corresponding text prompts (i.e., [EOS] token). Despite their good performance, we find that they discard the rich patch-level semantic information inherent in CLIP's encoders. For instance, when recognizing a rabbit, local patch"},"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":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.13835","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-13T17:56:23Z","cross_cats_sorted":[],"title_canon_sha256":"bbe520d2b68ca1ed89b41ab3c8ced9b1702690ee969a576c1c0529a747da49d6","abstract_canon_sha256":"c26a1c0cad34eea2175be1335d6ebfaed5362e077320d5b34f78c4ef05e8ab4f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T02:44:14.945301Z","signature_b64":"F4mfkc4BYiSOF0bhBBaad9doArHcHCyS7dSIBX3fAylDZrKE84WTgdYnFmIpq606Miv9Q3KkE4GN5JtnFRgfCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a793a70fbc00b0779429b00822bc92e3e181b1c1a3ef2a907fb50e73c0b47cc2","last_reissued_at":"2026-05-18T02:44:14.944708Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T02:44:14.944708Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Unlocking Patch-Level Features for CLIP-Based Class-Incremental Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Aligning CLIP patch features to semantic descriptions improves class-incremental learning.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Da-Wei Zhou, Hao Sun, Zi-Jun Ding","submitted_at":"2026-05-13T17:56:23Z","abstract_excerpt":"Class-Incremental Learning (CIL) enables models to continuously integrate new knowledge while mitigating catastrophic forgetting. Driven by the remarkable generalization of CLIP, leveraging pre-trained vision-language models has become a dominant paradigm in CIL. However, current work primarily focuses on aligning global image embeddings (i.e., [CLS] token) with their corresponding text prompts (i.e., [EOS] token). Despite their good performance, we find that they discard the rich patch-level semantic information inherent in CLIP's encoders. For instance, when recognizing a rabbit, local patch"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Extensive experiments demonstrate that SPA achieves state-of-the-art performance.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That GPT-5-generated class-wise semantic descriptions reliably identify discriminative patches and that optimal transport alignment between selected patches and semantic tokens yields a meaningful recognition improvement beyond global embeddings.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SPA unlocks patch-level features in CLIP for class-incremental learning via semantic-guided selection and optimal transport alignment with class descriptions, plus projectors and pseudo-feature replay to reduce forgetting.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Aligning CLIP patch features to semantic descriptions improves class-incremental learning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"27bdb55f10c22e42961089c59148ccdacf9ac3bec2f98e727ce5917e37b6accd"},"source":{"id":"2605.13835","kind":"arxiv","version":1},"verdict":{"id":"248d67c7-6a84-4771-8a88-be4cb1b92b19","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:08:58.877996Z","strongest_claim":"Extensive experiments demonstrate that SPA achieves state-of-the-art performance.","one_line_summary":"SPA unlocks patch-level features in CLIP for class-incremental learning via semantic-guided selection and optimal transport alignment with class descriptions, plus projectors and pseudo-feature replay to reduce forgetting.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That GPT-5-generated class-wise semantic descriptions reliably identify discriminative patches and that optimal transport alignment between selected patches and semantic tokens yields a meaningful recognition improvement beyond global embeddings.","pith_extraction_headline":"Aligning CLIP patch features to semantic descriptions improves class-incremental learning."},"references":{"count":72,"sample":[{"doi":"","year":2018,"title":"Memory aware synapses: Learning what (not) to forget","work_id":"0925aaf0-e501-443e-a639-34de9b7a3620","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Qwen-vl: A versatile vision-language model for understanding, localization.Text Reading, and Beyond, 2(1):1, 2023","work_id":"3e704b76-920c-4ff8-85b2-7603f3c5ddd3","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models.Advances in neural information processing systems, 32, 2019","work_id":"1e1c2889-d7d1-4115-a84d-eb33cf223329","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2014,"title":"Food-101–mining discriminative components with random forests","work_id":"5fc3ab66-8c71-4e08-9abb-b01e46165194","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Efficient lifelong learning with A-GEM.CoRR, abs/1812.00420","work_id":"c1e3aa81-39b5-44a3-a1b5-6fa343e0b956","ref_index":5,"cited_arxiv_id":"1812.00420","is_internal_anchor":true}],"resolved_work":72,"snapshot_sha256":"b6bfcde75964b003cec5e88700e1db4610c5422d61988370c0d03a15bc560de8","internal_anchors":7},"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":"2605.13835","created_at":"2026-05-18T02:44:14.944793+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.13835v1","created_at":"2026-05-18T02:44:14.944793+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13835","created_at":"2026-05-18T02:44:14.944793+00:00"},{"alias_kind":"pith_short_12","alias_value":"U6J2OD54ACYH","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"U6J2OD54ACYHPFBJ","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"U6J2OD54","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/U6J2OD54ACYHPFBJWAECFPES4P","json":"https://pith.science/pith/U6J2OD54ACYHPFBJWAECFPES4P.json","graph_json":"https://pith.science/api/pith-number/U6J2OD54ACYHPFBJWAECFPES4P/graph.json","events_json":"https://pith.science/api/pith-number/U6J2OD54ACYHPFBJWAECFPES4P/events.json","paper":"https://pith.science/paper/U6J2OD54"},"agent_actions":{"view_html":"https://pith.science/pith/U6J2OD54ACYHPFBJWAECFPES4P","download_json":"https://pith.science/pith/U6J2OD54ACYHPFBJWAECFPES4P.json","view_paper":"https://pith.science/paper/U6J2OD54","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.13835&json=true","fetch_graph":"https://pith.science/api/pith-number/U6J2OD54ACYHPFBJWAECFPES4P/graph.json","fetch_events":"https://pith.science/api/pith-number/U6J2OD54ACYHPFBJWAECFPES4P/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/U6J2OD54ACYHPFBJWAECFPES4P/action/timestamp_anchor","attest_storage":"https://pith.science/pith/U6J2OD54ACYHPFBJWAECFPES4P/action/storage_attestation","attest_author":"https://pith.science/pith/U6J2OD54ACYHPFBJWAECFPES4P/action/author_attestation","sign_citation":"https://pith.science/pith/U6J2OD54ACYHPFBJWAECFPES4P/action/citation_signature","submit_replication":"https://pith.science/pith/U6J2OD54ACYHPFBJWAECFPES4P/action/replication_record"}},"created_at":"2026-05-18T02:44:14.944793+00:00","updated_at":"2026-05-18T02:44:14.944793+00:00"}