{"id":"W3213180921","doi":"10.18653/v1/2021.emnlp-main.603","title":"Universal-KD: Attention-based Output-Grounded Intermediate Layer Knowledge Distillation","year":2021,"lang":"en","type":"article","venue":"Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing","topic":"Topic Modeling","field":"Computer Science","cited_by":18,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Waterloo","funders":"","keywords":"Computer science; Layer (electronics); Distillation; Matching (statistics); Interpretability; Projection (relational algebra); Base (topology); Space (punctuation); Architecture; Artificial intelligence; Deep learning; Computer architecture; Machine learning; Algorithm; Mathematics; Nanotechnology; Chromatography; Operating system; Materials science; Chemistry","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":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008527695,0.0002367561,0.0003583735,0.0001908473,0.0001290096,0.0002700357,0.001021373,0.0001411289,0.00001928073],"category_scores_gemma":[0.001398537,0.0001768341,0.0001441482,0.001190857,0.0001036087,0.0005192872,0.0004914308,0.0006171339,0.000003437983],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001813978,"about_ca_system_score_gemma":0.0003234548,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001192993,"about_ca_topic_score_gemma":0.00001257149,"domain_scores_codex":[0.9979978,0.0001611217,0.0004650662,0.0006452379,0.0003916377,0.0003391556],"domain_scores_gemma":[0.9983684,0.0002990369,0.0003243437,0.0002715064,0.0006595681,0.00007716658],"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":[0.00004430267,0.0001948558,0.005383664,0.0004079418,0.00002625667,0.00001198843,0.00546249,0.00008385377,0.06636719,0.01196829,0.00003621514,0.910013],"study_design_scores_gemma":[0.0007080266,0.00003823698,0.01104244,0.001090848,0.00003204455,0.000009347238,0.001574158,0.9253258,0.05249839,0.007180464,0.0001146685,0.0003856334],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.3435652,0.003309703,0.6312937,0.006508523,0.00137135,0.0004401724,0.000004475752,0.0001927762,0.01331412],"genre_scores_gemma":[0.7555471,0.000004856316,0.2437306,0.0001727623,0.00006304575,0.000007893663,0.000001577716,0.00001124146,0.0004609142],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9252419,"threshold_uncertainty_score":0.7211084,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.07139261601228206,"score_gpt":0.3974343703737435,"score_spread":0.3260417543614614,"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."}}