{"id":"W4385573095","doi":"10.18653/v1/2022.emnlp-main.278","title":"Mixture of Attention Heads: Selecting Attention Heads Per Token","year":2022,"lang":"en","type":"article","venue":"","topic":"Advanced Graph Neural Networks","field":"Computer Science","cited_by":23,"is_retracted":false,"has_abstract":true,"ca_institutions":"Université de Montréal; Mila - Quebec Artificial Intelligence Institute","funders":"State Key Laboratory of Software Development Environment","keywords":"Computer science; Interpretability; Feed forward; Artificial intelligence; Security token; Computation; Transformer; Set (abstract data type); Machine learning; Computer network; Algorithm","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":[0.001679837,0.001287476,0.00112414,0.0007442652,0.0005914907,0.0009467464,0.002447383,0.001576026,0.004493019],"category_scores_gemma":[0.004441857,0.0008529482,0.0009450494,0.0007636106,0.001038048,0.003232909,0.002629313,0.002113592,0.001219205],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001055085,"about_ca_system_score_gemma":0.00134251,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.007351493,"about_ca_topic_score_gemma":0.01277011,"domain_scores_codex":[0.9993972,0.0001958104,0.00002371662,0.0001817894,0.00007700715,0.0001245814],"domain_scores_gemma":[0.998832,0.0005774473,0.00007433654,0.0002295188,0.0001860349,0.0001006994],"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":[0.001082441,0.0002932647,0.004885954,0.0002004136,0.0002502231,0.0002880696,0.0005063572,0.2606744,0.02922123,0.03219047,0.01171907,0.6586881],"study_design_scores_gemma":[0.0000303303,0.00006464346,0.0003939894,0.00001119583,0.00005080698,0.00006622309,0.00003477698,0.9711607,0.008269472,0.0183974,0.001497392,0.00002312095],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.03382946,0.0003091315,0.9597239,0.0003079806,0.00008572377,0.0001115725,0.000106193,0.003550227,0.001975818],"genre_scores_gemma":[0.7060053,0.000247875,0.2831578,0.0006301284,0.00009551376,0.0002008161,0.0003382132,0.0004032566,0.008920939],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.007351493,"threshold_uncertainty_score":0.01503068,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01061476401515083,"score_gpt":0.2437902451232547,"score_spread":0.2331754811081039,"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."}}