{"id":822442,"date":"2022-02-25T05:36:47","date_gmt":"2022-02-25T13:36:47","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-research-item&#038;p=822442"},"modified":"2022-02-25T05:37:37","modified_gmt":"2022-02-25T13:37:37","slug":"investigating-the-role-of-negatives-in-contrastive-representation-learning","status":"publish","type":"msr-research-item","link":"https:\/\/www.microsoft.com\/en-us\/research\/publication\/investigating-the-role-of-negatives-in-contrastive-representation-learning\/","title":{"rendered":"Investigating the Role of Negatives in Contrastive Representation Learning"},"content":{"rendered":"<p><span dir=\"ltr\" role=\"presentation\">Noise contrastive learning is a popular tech<\/span><span dir=\"ltr\" role=\"presentation\">nique for unsupervised representation learn<\/span><span dir=\"ltr\" role=\"presentation\">ing.<\/span> <span dir=\"ltr\" role=\"presentation\">In this approach, a representation is <\/span><span dir=\"ltr\" role=\"presentation\">obtained via reduction to supervised learning, <\/span><span dir=\"ltr\" role=\"presentation\">where given a notion of semantic similarity, <\/span><span dir=\"ltr\" role=\"presentation\">the learner tries to distinguish a similar (pos<\/span><span dir=\"ltr\" role=\"presentation\">itive) example from a collection of random <\/span><span dir=\"ltr\" role=\"presentation\">(negative) examples. The success of modern <\/span><span dir=\"ltr\" role=\"presentation\">contrastive learning pipelines relies on many <\/span><span dir=\"ltr\" role=\"presentation\">design decisions, such as the choice of data <\/span><span dir=\"ltr\" role=\"presentation\">augmentation, the number of negative exam<\/span><span dir=\"ltr\" role=\"presentation\">ples, and the batch size; however, there is <\/span><span dir=\"ltr\" role=\"presentation\">limited understanding as to how these param<\/span><span dir=\"ltr\" role=\"presentation\">eters interact and affect downstream perfor<\/span><span dir=\"ltr\" role=\"presentation\">mance. We focus on disambiguating the role <\/span><span dir=\"ltr\" role=\"presentation\">of one of these parameters: the number of <\/span><span dir=\"ltr\" role=\"presentation\">negative examples.<\/span> <span dir=\"ltr\" role=\"presentation\">Theoretically, we show <\/span><span dir=\"ltr\" role=\"presentation\">the existence of a collision-coverage trade-off <\/span><span dir=\"ltr\" role=\"presentation\">suggesting that the optimal number of nega<\/span><span dir=\"ltr\" role=\"presentation\">tive examples should scale with the number <\/span><span dir=\"ltr\" role=\"presentation\">of underlying concepts in the data. Empiri<\/span><span dir=\"ltr\" role=\"presentation\">cally, we scrutinize the role of the number of <\/span><span dir=\"ltr\" role=\"presentation\">negatives in both NLP and vision tasks.<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Noise contrastive learning is a popular technique for unsupervised representation learning. In this approach, a representation is obtained via reduction to supervised learning, where given a notion of semantic similarity, the learner tries to distinguish a similar (positive) example from a collection of random (negative) examples. The success of modern contrastive learning pipelines relies on [&hellip;]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"footnotes":""},"msr-content-type":[3],"msr-research-highlight":[],"research-area":[13556],"msr-publication-type":[193716],"msr-product-type":[],"msr-focus-area":[],"msr-platform":[],"msr-download-source":[],"msr-locale":[268875],"msr-field-of-study":[],"msr-conference":[260149],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-822442","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-artificial-intelligence","msr-locale-en_us"],"msr_publishername":"","msr_edition":"","msr_affiliation":"","msr_published_date":"2022-2-1","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"","msr_volume":"","msr_number":"","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"","msr_how_published":"","msr_notes":"","msr_highlight_text":"","msr_release_tracker_id":"","msr_original_fields_of_study":"","msr_download_urls":"","msr_external_url":"","msr_secondary_video_url":"","msr_longbiography":"","msr_microsoftintellectualproperty":1,"msr_main_download":"","msr_publicationurl":"","msr_doi":"","msr_publication_uploader":[{"type":"url","viewUrl":"false","id":"false","title":"https:\/\/arxiv.org\/abs\/2106.09943","label_id":"243109","label":0}],"msr_related_uploader":"","msr_attachments":[],"msr-author-ordering":[{"type":"user_nicename","value":"Jordan Ash","user_id":39826,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Jordan Ash"},{"type":"user_nicename","value":"Surbhi Goel","user_id":39543,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Surbhi Goel"},{"type":"user_nicename","value":"Akshay Krishnamurthy","user_id":30913,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Akshay Krishnamurthy"},{"type":"user_nicename","value":"Dipendra Misra","user_id":38607,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Dipendra Misra"}],"msr_impact_theme":[],"msr_research_lab":[],"msr_event":[821218],"msr_group":[144902],"msr_project":[],"publication":[],"video":[],"download":[],"msr_publication_type":"inproceedings","_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/822442"}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-research-item"}],"version-history":[{"count":1,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/822442\/revisions"}],"predecessor-version":[{"id":822445,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/822442\/revisions\/822445"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=822442"}],"wp:term":[{"taxonomy":"msr-content-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-content-type?post=822442"},{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=822442"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=822442"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=822442"},{"taxonomy":"msr-product-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-product-type?post=822442"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=822442"},{"taxonomy":"msr-platform","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-platform?post=822442"},{"taxonomy":"msr-download-source","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-download-source?post=822442"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=822442"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=822442"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=822442"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=822442"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=822442"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=822442"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}