{"id":"W6966481125","doi":"10.48448/fyjq-3w14","title":"Preserving Privacy Through Dememorization: An Unlearning Technique For Mitigating Memorization Risks In Language Models","year":2023,"lang":"en","type":"other","venue":"Open MIND","topic":"","field":"","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Windsor; Artificial Intelligence in Medicine (Canada)","funders":"","keywords":"Memorization; Language model; Reinforcement learning; Replication (statistics); Similarity (geometry); Generative grammar; Information privacy","routes":{"ca_aff":true,"ca_fund":false,"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":[0.002948091,0.001043996,0.0007459824,0.0006257273,0.0006378356,0.001323319,0.002243635,0.001508843,0.003681096],"category_scores_gemma":[0.01930906,0.0004231881,0.000815596,0.0005885274,0.002143467,0.005327907,0.00426271,0.003222368,0.001489037],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007880699,"about_ca_system_score_gemma":0.001346659,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001428983,"about_ca_topic_score_gemma":0.002627296,"domain_scores_codex":[0.997776,0.0009963443,0.0001002424,0.0005265898,0.000445898,0.0001548773],"domain_scores_gemma":[0.9907623,0.004711758,0.0006742259,0.003247631,0.0003816406,0.0002224708],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0006906055,0.0005691487,0.003578913,0.0003565594,0.0001736085,0.0005636354,0.001022946,0.2476598,0.01559731,0.08501218,0.01549994,0.6292754],"study_design_scores_gemma":[0.00003910226,0.0001096847,0.0002533122,0.00003334612,0.00002326554,0.0002272097,0.00005582178,0.8968132,0.01115648,0.08726965,0.00398953,0.00002942757],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.03316825,0.0005215101,0.9569377,0.0009548955,0.00007630906,0.00009047669,0.0001987408,0.00413649,0.003915663],"genre_scores_gemma":[0.7777274,0.0004491053,0.2108463,0.0009649318,0.0001222794,0.0002192997,0.0005406318,0.0006593527,0.008470783],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.003681096,"threshold_uncertainty_score":0.0155912,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.1358138545269244,"score_gpt":0.408978155749854,"score_spread":0.2731643012229296,"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."}}