{"id":"W4389523936","doi":"10.18653/v1/2023.emnlp-main.265","title":"Preserving Privacy Through Dememorization: An Unlearning Technique For Mitigating Memorization Risks In Language Models","year":2023,"lang":"en","type":"article","venue":"","topic":"Privacy-Preserving Technologies in Data","field":"Computer Science","cited_by":10,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Windsor","funders":"Natural Sciences and Engineering Research Council of Canada; Alliance de recherche numérique du Canada","keywords":"Memorization; Computer science; Reinforcement learning; Replication (statistics); Artificial intelligence; Similarity (geometry); Quality (philosophy); Cognitive psychology; Psychology","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004431078,0.0008945082,0.0008962452,0.0008341279,0.001103751,0.001895033,0.00240402,0.001499332,0.001835],"category_scores_gemma":[0.03825588,0.0005426353,0.001160777,0.0007191553,0.003098322,0.01029726,0.00485252,0.004750567,0.0003646535],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007899337,"about_ca_system_score_gemma":0.001419638,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0006844273,"about_ca_topic_score_gemma":0.0008604369,"domain_scores_codex":[0.9949287,0.002059295,0.0003958615,0.0009384993,0.001224969,0.0004526195],"domain_scores_gemma":[0.9614814,0.01893527,0.002870094,0.01438299,0.001876567,0.0004536272],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","study_design_scores_codex":[0.00124093,0.0007907468,0.008455803,0.0005129124,0.0002982271,0.0006566365,0.002761768,0.1499621,0.04147882,0.2511466,0.005588238,0.5371072],"study_design_scores_gemma":[0.00004301746,0.0002473578,0.0005005809,0.00006179773,0.0001046739,0.0004580234,0.0002654755,0.6967523,0.04716061,0.2511137,0.00322322,0.00006924292],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.04379239,0.000190661,0.9529121,0.0008283706,0.00005344067,0.00008154123,0.00006162092,0.0009139421,0.00116599],"genre_scores_gemma":[0.8469029,0.0001955704,0.1498225,0.000493528,0.0001147357,0.0001337576,0.00009914146,0.0001360081,0.002101821],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.004431078,"threshold_uncertainty_score":0.0234341,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.09904560295695426,"score_gpt":0.3578418422760816,"score_spread":0.2587962393191273,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}