{"id":944886,"date":"2023-05-30T08:43:36","date_gmt":"2023-05-30T15:43:36","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/"},"modified":"2024-06-18T09:36:46","modified_gmt":"2024-06-18T16:36:46","slug":"selective-pre-training-for-private-fine-tuning","status":"publish","type":"msr-research-item","link":"https:\/\/www.microsoft.com\/en-us\/research\/publication\/selective-pre-training-for-private-fine-tuning\/","title":{"rendered":"Selective Pre-training for Private Fine-tuning"},"content":{"rendered":"

Suppose we want to train text prediction models in email clients or word processors. The models must preserve the privacy of user data and adhere to a specific fixed size to meet memory and inference time requirements. We introduce a generic framework to solve this problem. Specifically, we are given a public dataset\u00a0D_<\/span>pub<\/span><\/span><\/span><\/span><\/span>\u00a0and a private dataset\u00a0D_<\/span>priv<\/span><\/span><\/span><\/span><\/span>\u00a0corresponding to a downstream task\u00a0T<\/span><\/span><\/span><\/span>. How should we pre-train a fixed-size model\u00a0M<\/span><\/span><\/span><\/span>\u00a0on\u00a0D_<\/span>pub<\/span><\/span><\/span><\/span><\/span>\u00a0and fine-tune it on\u00a0D_<\/span>priv<\/span><\/span><\/span><\/span><\/span>\u00a0such that performance of\u00a0M<\/span><\/span><\/span><\/span>\u00a0with respect to\u00a0T<\/span><\/span><\/span><\/span>\u00a0is maximized and\u00a0M<\/span><\/span><\/span><\/span>\u00a0satisfies differential privacy with respect to\u00a0D_<\/span>priv<\/span><\/span><\/span><\/span><\/span>? We show that pre-training on a subset of dataset D_<\/span>pub<\/span><\/span><\/span><\/span><\/span>\u00a0that brings the public distribution closer to the private distribution is a crucial ingredient to maximize the transfer learning abilities of\u00a0M<\/span><\/span><\/span><\/span> after pre-training, especially in the regimes where model sizes are relatively small. Besides performance improvements, our framework also shows that with careful pre-training and private fine-tuning, smaller models can match the performance of much larger models, highlighting the promise of differentially private training as a tool for model compression and efficiency.<\/p>\n","protected":false},"excerpt":{"rendered":"

Suppose we want to train text prediction models in email clients or word processors. The models must preserve the privacy of user data and adhere to a specific fixed size to meet memory and inference time requirements. We introduce a generic framework to solve this problem. Specifically, we are given a public dataset\u00a0D_pub\u00a0and a private […]<\/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":[13561,13558],"msr-publication-type":[193715],"msr-product-type":[],"msr-focus-area":[],"msr-platform":[],"msr-download-source":[],"msr-locale":[268875],"msr-field-of-study":[246694,251401,253561],"msr-conference":[],"msr-journal":[],"msr-impact-theme":[264846],"msr-pillar":[],"class_list":["post-944886","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-algorithms","msr-research-area-security-privacy-cryptography","msr-locale-en_us","msr-field-of-study-artificial-intelligence","msr-field-of-study-data-privacy","msr-field-of-study-differential-privacy"],"msr_publishername":"","msr_edition":"","msr_affiliation":"","msr_published_date":"2024-5-1","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"TMLR","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\/2305.13865","label_id":"243109","label":0},{"type":"url","viewUrl":"false","id":"false","title":"https:\/\/openreview.net\/forum?id=y3u8OpPHxz","label_id":"243109","label":0}],"msr_related_uploader":"","msr_attachments":[],"msr-author-ordering":[{"type":"guest","value":"da-yu","user_id":783421,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=da-yu"},{"type":"user_nicename","value":"Sivakanth Gopi","user_id":37830,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Sivakanth Gopi"},{"type":"user_nicename","value":"Janardhan (Jana) Kulkarni","user_id":32147,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Janardhan (Jana) Kulkarni"},{"type":"user_nicename","value":"Zinan Lin","user_id":42327,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Zinan Lin"},{"type":"text","value":"Saurabh Naik","user_id":0,"rest_url":false},{"type":"text","value":"Tomasz Lukasz Religa","user_id":0,"rest_url":false},{"type":"text","value":"Jian Yin","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Huishuai Zhang","user_id":37781,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Huishuai Zhang"}],"msr_impact_theme":["Computing foundations"],"msr_research_lab":[199560,199565],"msr_event":[],"msr_group":[437022,761911,793670,802999],"msr_project":[556311],"publication":[],"video":[],"download":[],"msr_publication_type":"article","_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/944886"}],"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":2,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/944886\/revisions"}],"predecessor-version":[{"id":1048383,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/944886\/revisions\/1048383"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=944886"}],"wp:term":[{"taxonomy":"msr-content-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-content-type?post=944886"},{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=944886"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=944886"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=944886"},{"taxonomy":"msr-product-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-product-type?post=944886"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=944886"},{"taxonomy":"msr-platform","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-platform?post=944886"},{"taxonomy":"msr-download-source","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-download-source?post=944886"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=944886"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=944886"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=944886"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=944886"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=944886"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=944886"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}