{"paper":{"title":"Class Unlearning via Depth-Aware Removal of Forget-Specific Directions","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"By projecting out forget-specific directions layer by layer with depth-aware scaling, a closed-form method achieves class unlearning closer to full retraining than prior approaches.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CV","authors_text":"Arman Hatami, Ilya E. Monosov, Romina Aalishah","submitted_at":"2026-04-16T15:46:02Z","abstract_excerpt":"Machine unlearning aims to remove targeted knowledge from a trained model without the cost of retraining from scratch. In class unlearning, however, reducing accuracy on forget classes does not necessarily imply true forgetting: forgotten information can remain encoded in internal representations, and apparent forgetting may arise from classifier-head suppression rather than representational removal. We show that existing class-unlearning methods often exhibit weak or negative selectivity, preserve forget-class structure in deep representations, or rely heavily on final-layer bias shifts. We t"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across MNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet, and across convolutional and transformer architectures, DAMP more closely resembles the retraining gold standard than some of the prior methods, improving selective forgetting while better preserving retain-class performance and reducing residual forget-class structure in deep layers.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That forget directions extracted as residuals relative to retain-class prototypes at each layer accurately isolate the targeted knowledge, and that the parameter-free depth-aware scaling derived from probe separability will preserve utility without introducing new failure modes on retain classes.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DAMP performs one-shot class unlearning by extracting and projecting out forget-specific residual directions at each network depth using class prototypes and a separability-derived scaling rule.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"By projecting out forget-specific directions layer by layer with depth-aware scaling, a closed-form method achieves class unlearning closer to full retraining than prior approaches.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b40c76d111c9bac4a0b4b8059e503eca22e0e488a746da4a466244dc9022aee2"},"source":{"id":"2604.15166","kind":"arxiv","version":2},"verdict":{"id":"ce474362-0a7f-4108-bf0c-ce88d88df969","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T11:42:48.804913Z","strongest_claim":"Across MNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet, and across convolutional and transformer architectures, DAMP more closely resembles the retraining gold standard than some of the prior methods, improving selective forgetting while better preserving retain-class performance and reducing residual forget-class structure in deep layers.","one_line_summary":"DAMP performs one-shot class unlearning by extracting and projecting out forget-specific residual directions at each network depth using class prototypes and a separability-derived scaling rule.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That forget directions extracted as residuals relative to retain-class prototypes at each layer accurately isolate the targeted knowledge, and that the parameter-free depth-aware scaling derived from probe separability will preserve utility without introducing new failure modes on retain classes.","pith_extraction_headline":"By projecting out forget-specific directions layer by layer with depth-aware scaling, a closed-form method achieves class unlearning closer to full retraining than prior approaches."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.15166/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"}