{"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":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002675343,0.0001234667,0.0001550164,0.0001235032,0.0003109337,0.00003709115,0.0005229161,0.00003693656,0.0001112555],"category_scores_gemma":[0.00001103972,0.0001204628,0.0001258592,0.0008085847,0.00002270477,0.0004904807,0.000424379,0.0002919068,0.000009898708],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000617976,"about_ca_system_score_gemma":0.00002007611,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000236604,"about_ca_topic_score_gemma":0.00001407842,"domain_scores_codex":[0.9985128,0.0001334393,0.0002745615,0.000390367,0.0003967779,0.0002920503],"domain_scores_gemma":[0.9992905,0.00005706841,0.0001493131,0.0003679598,0.00008237789,0.00005273629],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"bench_or_experimental","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0001133729,0.0008473187,0.09368031,0.0001576552,0.0001695463,0.00005334501,0.001843484,0.05333063,0.4137343,0.118707,0.02731738,0.2900457],"study_design_scores_gemma":[0.003543235,0.002383382,0.1913642,0.0001300688,0.00008182104,0.0006437765,0.001337337,0.7182934,0.01648445,0.0335074,0.029951,0.002279855],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.3573837,0.0003183227,0.6381722,0.001667408,0.0007610937,0.0002648122,0.000002451881,0.0003002729,0.001129735],"genre_scores_gemma":[0.9713095,0.00001111013,0.02695978,0.0004288734,0.0000628001,0.00002606537,0.00001080437,0.00001158854,0.00117949],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.6649628,"threshold_uncertainty_score":0.4912329,"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."}}