{"id":"W4382599796","doi":"10.3389/fcomp.2023.1178450","title":"Self-attention in vision transformers performs perceptual grouping, not attention","year":2023,"lang":"en","type":"article","venue":"Frontiers in Computer Science","topic":"Visual Attention and Saliency Detection","field":"Computer Science","cited_by":37,"is_retracted":false,"has_abstract":true,"ca_institutions":"York University","funders":"Air Force Office of Scientific Research; Natural Sciences and Engineering Research Council of Canada; Canada Research Chairs","keywords":"Artificial intelligence; Salience (neuroscience); Visual perception; Computer science; Visual attention; Computational model; Transformer; Perception; Salient; Cognitive psychology; Psychology; Neuroscience; Engineering","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":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002207885,0.0002299867,0.0002550338,0.001995217,0.0003752279,0.0004215198,0.001295794,0.0001090477,0.000004989079],"category_scores_gemma":[0.00001709227,0.0002225697,0.0001247297,0.005753106,0.0002204891,0.003239603,0.0003153979,0.0003081082,0.0001220409],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004217189,"about_ca_system_score_gemma":0.0001146177,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003147977,"about_ca_topic_score_gemma":0.0000256291,"domain_scores_codex":[0.99652,0.0001099582,0.0005421466,0.001049872,0.0009946388,0.000783357],"domain_scores_gemma":[0.9992136,0.00002198311,0.0001044281,0.00041156,0.00009622548,0.0001521637],"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":[0.00003995574,0.000492854,0.1292867,0.000119955,0.00001351433,0.00007009759,0.006094865,0.003484265,0.00842799,0.002266973,0.001737696,0.8479652],"study_design_scores_gemma":[0.0005158049,0.000175602,0.3295758,0.00005134778,0.000001647489,0.00001046899,0.0001774485,0.6685559,0.0001473705,0.0004262992,0.0001515469,0.0002107587],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.4288372,0.00000883557,0.5666344,0.0004713951,0.003400955,0.0002204301,5.164717e-7,0.0003173014,0.0001090433],"genre_scores_gemma":[0.9236105,0.00007239688,0.07592269,0.0001917575,0.00007398258,0.00002689683,0.000005628046,0.00001126933,0.00008487637],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8477544,"threshold_uncertainty_score":0.9076129,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01226093722006374,"score_gpt":0.2660569458484697,"score_spread":0.253796008628406,"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."}}