Publikation: Differentiable Top-k Classification Learning
Dateien
Datum
Autor:innen
Herausgeber:innen
ISSN der Zeitschrift
Electronic ISSN
ISBN
Bibliografische Daten
Verlag
Schriftenreihe
Auflagebezeichnung
URI (zitierfähiger Link)
Internationale Patentnummer
Link zur Lizenz
Angaben zur Forschungsförderung
Projekt
Open Access-Veröffentlichung
Core Facility der Universität Konstanz
Titel in einer weiteren Sprache
Publikationstyp
Publikationsstatus
Erschienen in
Zusammenfassung
The top-k classification accuracy is one of the core metrics in machine learning. Here, k is conventionally a positive integer, such as 1 or 5, leading to top-1 or top-5 training objectives. In this work, we relax this assumption and optimize the model for multiple k simultaneously instead of using a single k. Leveraging recent advances in differentiable sorting and ranking, we propose a family of differentiable top-k cross-entropy classification losses. This allows training while not only considering the top-1 prediction, but also, e.g., the top-2 and top-5 predictions. We evaluate the proposed losses for fine-tuning on state-of-the-art architectures, as well as for training from scratch. We find that relaxing k not only produces better top-5 accuracies, but also leads to top-1 accuracy improvements. When fine-tuning publicly available ImageNet models, we achieve a new state-of-the-art for these models.
Zusammenfassung in einer weiteren Sprache
Fachgebiet (DDC)
Schlagwörter
Konferenz
Rezension
Zitieren
ISO 690
PETERSEN, Felix, Hilde KUEHNE, Christian BORGELT, Oliver DEUSSEN, 2022. Differentiable Top-k Classification Learning. 39th International Conference on Machine Learning : PLMR 162. Baltimore, Maryland, 17. Juli 2022 - 23. Juli 2022. In: CHAUDHURI, Kamalika, ed., Stefanie JEGELKA, ed., Le SONG, ed. and others. International Conference on Machine Learning, Vol. 162. PLMR, 2022, pp. 17656-17668BibTex
@inproceedings{Petersen2022Diffe-67074, year={2022}, title={Differentiable Top-k Classification Learning}, url={https://proceedings.mlr.press/v162/petersen22a.html}, publisher={PLMR}, booktitle={International Conference on Machine Learning, Vol. 162}, pages={17656--17668}, editor={Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le}, author={Petersen, Felix and Kuehne, Hilde and Borgelt, Christian and Deussen, Oliver} }
RDF
<rdf:RDF xmlns:dcterms="http://purl.org/dc/terms/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#" xmlns:bibo="http://purl.org/ontology/bibo/" xmlns:dspace="http://digital-repositories.org/ontologies/dspace/0.1.0#" xmlns:foaf="http://xmlns.com/foaf/0.1/" xmlns:void="http://rdfs.org/ns/void#" xmlns:xsd="http://www.w3.org/2001/XMLSchema#" > <rdf:Description rdf:about="https://kops.uni-konstanz.de/server/rdf/resource/123456789/67074"> <dc:creator>Kuehne, Hilde</dc:creator> <dcterms:rights rdf:resource="https://rightsstatements.org/page/InC/1.0/"/> <dc:contributor>Petersen, Felix</dc:contributor> <dc:creator>Borgelt, Christian</dc:creator> <dc:creator>Petersen, Felix</dc:creator> <dc:rights>terms-of-use</dc:rights> <dspace:isPartOfCollection rdf:resource="https://kops.uni-konstanz.de/server/rdf/resource/123456789/36"/> <dc:contributor>Borgelt, Christian</dc:contributor> <dc:contributor>Deussen, Oliver</dc:contributor> <dcterms:available rdf:datatype="http://www.w3.org/2001/XMLSchema#dateTime">2023-06-07T07:14:00Z</dcterms:available> <dcterms:hasPart rdf:resource="https://kops.uni-konstanz.de/bitstream/123456789/67074/1/Petersen_2-1b1lxhiqw7rwh1.pdf"/> <dcterms:abstract>The top-k classification accuracy is one of the core metrics in machine learning. Here, k is conventionally a positive integer, such as 1 or 5, leading to top-1 or top-5 training objectives. In this work, we relax this assumption and optimize the model for multiple k simultaneously instead of using a single k. Leveraging recent advances in differentiable sorting and ranking, we propose a family of differentiable top-k cross-entropy classification losses. This allows training while not only considering the top-1 prediction, but also, e.g., the top-2 and top-5 predictions. We evaluate the proposed losses for fine-tuning on state-of-the-art architectures, as well as for training from scratch. We find that relaxing k not only produces better top-5 accuracies, but also leads to top-1 accuracy improvements. When fine-tuning publicly available ImageNet models, we achieve a new state-of-the-art for these models.</dcterms:abstract> <dcterms:title>Differentiable Top-k Classification Learning</dcterms:title> <dc:creator>Deussen, Oliver</dc:creator> <void:sparqlEndpoint rdf:resource="http://localhost/fuseki/dspace/sparql"/> <bibo:uri rdf:resource="https://kops.uni-konstanz.de/handle/123456789/67074"/> <dc:contributor>Kuehne, Hilde</dc:contributor> <dc:language>eng</dc:language> <dcterms:isPartOf rdf:resource="https://kops.uni-konstanz.de/server/rdf/resource/123456789/36"/> <dcterms:issued>2022</dcterms:issued> <dspace:hasBitstream rdf:resource="https://kops.uni-konstanz.de/bitstream/123456789/67074/1/Petersen_2-1b1lxhiqw7rwh1.pdf"/> <foaf:homepage rdf:resource="http://localhost:8080/"/> <dc:date rdf:datatype="http://www.w3.org/2001/XMLSchema#dateTime">2023-06-07T07:14:00Z</dc:date> </rdf:Description> </rdf:RDF>