Natural Analysts in Adaptive Data Analysis Tijana Zrnic joint with - PowerPoint PPT Presentation

Natural Analysts in Adaptive Data Analysis Tijana Zrnic 
 joint with Moritz Hardt

Adaptivity in Machine Learning

Adaptivity in Machine Learning data analyst 
 with training data

Adaptivity in Machine Learning model data analyst 
 with training data

Adaptivity in Machine Learning model test data set data analyst 
 with training data

Adaptivity in Machine Learning model test data set data analyst 
 with training data accuracy

Adaptivity in Machine Learning model tune hyperparams... test data set data analyst 
 with training data accuracy

Adaptivity in Machine Learning model test data set data analyst 
 with training data accuracy

Adaptivity in Machine Learning model test data set data analyst 
 with training data accuracy

Adaptivity in Machine Learning model tune hyperparams... test data set data analyst 
 with training data accuracy

Adaptivity in Machine Learning model test data set data analyst 
 with training data accuracy

Adaptivity in Machine Learning model test data set data analyst 
 with training data accuracy

Adaptivity in Machine Learning model tune hyperparams... test data set data analyst 
 with training data accuracy

<latexit sha1_base64="btWuKJH9/rCxCKL5tGKBdwWU5A=">AB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mqoMeiF48t2FpoQ9lsN+3azSbsToQS+gu8eFDEqz/Jm/GbZuDtj4YeLw3w8y8IJHCoOt+O4W19Y3NreJ2aWd3b/+gfHjUNnGqGW+xWMa6E1DpVC8hQIl7ySa0yiQ/CEY3878hyeujYjVPU4S7kd0qEQoGEUrNbFfrhVdw6ySrycVCBHo1/+6g1ilkZcIZPUmK7nJuhnVKNgk9LvdTwhLIxHfKupYpG3PjZ/NApObPKgISxtqWQzNXfExmNjJlEge2MKI7MsjcT/O6KYbXfiZUkiJXbLEoTCXBmMy+JgOhOUM5sYQyLeythI2opgxtNiUbgrf8ip16reRbXWvKzUb/I4inACp3AOHlxBHe6gAS1gwOEZXuHNeXRenHfnY9FacPKZY/gD5/MH4XeM/A=</latexit> Adaptivity in Machine Learning • After tested models, how well does the final model t generalize? 
 ‣ Depends on how the accuracies are computed

Classical Holdout vs Response Mechanism model data analyst 
 with training data accuracy test data set

Classical Holdout vs Response Mechanism model data analyst 
 with training data accuracy test data set • Reporting exact sample accuracy allows for great overfitting

Classical Holdout vs Response Mechanism model response mechanism 
 for reporting accuracy data analyst 
 with training data accuracy test data set • Reporting exact sample accuracy allows for great overfitting • Better bounds can be obtained by having a non-trivial response mechanism 
 in charge of reporting accuracy on the test data

Main Questions model data analyst 
 with training data accuracy test data set

Main Questions model response mechanism 
 for reporting accuracy data analyst 
 with training data accuracy test data set

Main Questions model response mechanism 
 for reporting accuracy data analyst 
 with training data accuracy test data set How do we construct a mechanism such that its responses generalize to the • population?

Main Questions model response mechanism 
 for reporting accuracy data analyst 
 with training data accuracy test data set How do we construct a mechanism such that its responses generalize to the • population? want 95% reported accuracy on test data ≈ 95% accuracy on fresh data from same population 
 ‣

Main Questions model response mechanism 
 for reporting accuracy data analyst 
 with training data accuracy test data set How do we construct a mechanism such that its responses generalize to the • population? want 95% reported accuracy on test data ≈ 95% accuracy on fresh data from same population 
 ‣ For such a good mechanism, how much does a possibly adversarial analyst overfit? •

Abstraction via Adaptive Data Analysis

Abstraction via Adaptive Data Analysis Framework of Dwork et al. (2015)

Abstraction via Adaptive Data Analysis Framework of Dwork et al. (2015) analyst

Abstraction via Adaptive Data Analysis Framework of Dwork et al. (2015) analyst mechanism

<latexit sha1_base64="uAzqYQ8pfQnZjL0XJEs1v308KOM=">AB83icbVDLSsNAFL2pr1pfVZduBovgqiRV0GXRjcuK9gFNKJPpB06mYR5CX0N9y4UMStP+POv3HSZqGtBwYO59zLPXPClDOlXfbKa2tb2xulbcrO7t7+wfVw6OSowktE0SnsheiBXlTNC2ZprTXiopjkNOu+HkNve7T1QqlohHPU1pEORYBEjWFvJ92OsxwTz7GFWGVRrbt2dA60SryA1KNAaVL/8YUJMTIUmHCvV9xUBxmWmhFOZxXfKJpiMsEj2rdU4JiqIJtnqEzqwxRlEj7hEZz9fdGhmOlpnFoJ/OMatnLxf+8vtHRdZAxkRpNBVkcigxHOkF5AWjIJCWaTy3BRDKbFZExlphoW1Negrf85VXSadS9i3rj/rLWvCnqKMJnMI5eHAFTbiDFrSBQArP8ApvjnFenHfnYzFacoqdY/gD5/MHxZeRgQ=</latexit> <latexit sha1_base64="iVYgqWYpJ7GNWJI9xnYwSPU6ZHk=">ACB3icbVDLSgMxFM34rPU16lKQYBFclZkq6LoxmVF+4DOWDJpg1NMkOSEcowOzf+ihsXirj1F9z5N2baQbT1QODknHu5954gZlRpx/myFhaXldWS2vl9Y3NrW17Z7elokRi0sQRi2QnQIowKkhTU81IJ5YE8YCRdjC6zP32PZGKRuJWj2PiczQNKQYaSP17AOPIz3EiKU3macohz/RnYnyj274lSdCeA8cQtSAQUaPfvT60c4URozJBSXdeJtZ8iqSlmJCt7iSIxwiM0IF1DBeJE+enkjgweGaUPw0iaJzScqL87UsSVGvPAVOZbqlkvF/zuokOz/2UijRODpoDBhUEcwDwX2qSRYs7EhCEtqdoV4iCTC2kSXh+DOnjxPWrWqe1KtXZ9W6hdFHCWwDw7BMXDBGaiDK9ATYDBA3gCL+DVerSerTfrfVq6YBU9e+APrI9vIBKZdQ=</latexit> <latexit sha1_base64="2QmInd+nHO60uhS7AYCvgpiU4a8=">AB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mqoMeiF48t2FpoQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm/GbZuDtj4YeLw3w8y8IBFcG9f9dgpr6xubW8Xt0s7u3v5B+fCoreNUMWyxWMSqE1CNgktsGW4EdhKFNAoEPgTj25n/8IRK81jem0mCfkSHkoecUWOlpuyXK27VnYOsEi8nFcjR6Je/eoOYpRFKwTVu5ifEzqgxnAqelXqoxoWxMh9i1VNItZ/ND52SM6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/7xuasJrP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2oszYbEo2BG/5VXSrlW9i2qteVmp3+RxFOETuEcPLiCOtxBA1rAOEZXuHNeXRenHfnY9FacPKZY/gD5/MH2F+M9g=</latexit> <latexit sha1_base64="GH4lnl7xiOBJEv5hdGjYXKo68Lw=">AB8nicbVDLSsNAFL3xWeur6tLNYBFclaQKuiy6cVnBPqANZTKdtEMnkzBzI5TQz3DjQhG3fo07/8ZJm4W2Hhg4nHMvc+4JEikMu63s7a+sbm1Xdop7+7tHxWjo7bJk414y0Wy1h3A2q4FIq3UKDk3URzGgWSd4LJXe53nrg2IlaPOE24H9GREqFgFK3U60cUx4zKrDkbVKpuzZ2DrBKvIFUo0BxUvrDmKURV8gkNabnuQn6GdUomOSzcj81PKFsQke8Z6miETd+No8I+dWGZIw1vYpJHP190ZGI2OmUWAn84hm2cvF/7xeiuGNnwmVpMgVW3wUpJgTPL7yVBozlBOLaFMC5uVsDHVlKFtqWxL8JZPXiXtes27rNUfrqN26KOEpzCGVyAB9fQgHtoQgsYxPAMr/DmoPivDsfi9E1p9g5gT9wPn8AiGeRag=</latexit> Abstraction via Adaptive Data Analysis Framework of Dwork et al. (2015) data set S analyst mechanism S ∼ P n P - population distribution - sample size n

<latexit sha1_base64="rslrDZ2B9h+NmotiRJ6kHYRkDtA=">AB6nicbVBNS8NAEJ34WetX1aOXxSJ4KkV9Fj04rGi/YA2lM120y7dbOLuRCihP8GLB0W8+ou8+W/ctjlo64OBx3szMwLEikMu63s7K6tr6xWdgqbu/s7u2XDg6bJk414w0Wy1i3A2q4FIo3UKDk7URzGgWSt4LRzdRvPXFtRKwecJxwP6IDJULBKFrp/rHn9Uplt+LOQJaJl5My5Kj3Sl/dfszSiCtkhrT8dwE/YxqFEzySbGbGp5QNqID3rFU0YgbP5udOiGnVumTMNa2FJKZ+nsio5Ex4yiwnRHFoVn0puJ/XifF8MrPhEpS5IrNF4WpJBiT6d+kLzRnKMeWUKaFvZWwIdWUoU2naEPwFl9eJs1qxTuvVO8uyrXrPI4CHMJnIEHl1CDW6hDAxgM4Ble4c2Rzovz7nzMW1ecfOYI/sD5/AECVI2d</latexit> <latexit sha1_base64="2QmInd+nHO60uhS7AYCvgpiU4a8=">AB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mqoMeiF48t2FpoQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm/GbZuDtj4YeLw3w8y8IBFcG9f9dgpr6xubW8Xt0s7u3v5B+fCoreNUMWyxWMSqE1CNgktsGW4EdhKFNAoEPgTj25n/8IRK81jem0mCfkSHkoecUWOlpuyXK27VnYOsEi8nFcjR6Je/eoOYpRFKwTVu5ifEzqgxnAqelXqoxoWxMh9i1VNItZ/ND52SM6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/7xuasJrP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2oszYbEo2BG/5VXSrlW9i2qteVmp3+RxFOETuEcPLiCOtxBA1rAOEZXuHNeXRenHfnY9FacPKZY/gD5/MH2F+M9g=</latexit> <latexit sha1_base64="uAzqYQ8pfQnZjL0XJEs1v308KOM=">AB83icbVDLSsNAFL2pr1pfVZduBovgqiRV0GXRjcuK9gFNKJPpB06mYR5CX0N9y4UMStP+POv3HSZqGtBwYO59zLPXPClDOlXfbKa2tb2xulbcrO7t7+wfVw6OSowktE0SnsheiBXlTNC2ZprTXiopjkNOu+HkNve7T1QqlohHPU1pEORYBEjWFvJ92OsxwTz7GFWGVRrbt2dA60SryA1KNAaVL/8YUJMTIUmHCvV9xUBxmWmhFOZxXfKJpiMsEj2rdU4JiqIJtnqEzqwxRlEj7hEZz9fdGhmOlpnFoJ/OMatnLxf+8vtHRdZAxkRpNBVkcigxHOkF5AWjIJCWaTy3BRDKbFZExlphoW1Negrf85VXSadS9i3rj/rLWvCnqKMJnMI5eHAFTbiDFrSBQArP8ApvjnFenHfnYzFacoqdY/gD5/MHxZeRgQ=</latexit> <latexit sha1_base64="iVYgqWYpJ7GNWJI9xnYwSPU6ZHk=">ACB3icbVDLSgMxFM34rPU16lKQYBFclZkq6LoxmVF+4DOWDJpg1NMkOSEcowOzf+ihsXirj1F9z5N2baQbT1QODknHu5954gZlRpx/myFhaXldWS2vl9Y3NrW17Z7elokRi0sQRi2QnQIowKkhTU81IJ5YE8YCRdjC6zP32PZGKRuJWj2PiczQNKQYaSP17AOPIz3EiKU3macohz/RnYnyj274lSdCeA8cQtSAQUaPfvT60c4URozJBSXdeJtZ8iqSlmJCt7iSIxwiM0IF1DBeJE+enkjgweGaUPw0iaJzScqL87UsSVGvPAVOZbqlkvF/zuokOz/2UijRODpoDBhUEcwDwX2qSRYs7EhCEtqdoV4iCTC2kSXh+DOnjxPWrWqe1KtXZ9W6hdFHCWwDw7BMXDBGaiDK9ATYDBA3gCL+DVerSerTfrfVq6YBU9e+APrI9vIBKZdQ=</latexit> <latexit sha1_base64="GH4lnl7xiOBJEv5hdGjYXKo68Lw=">AB8nicbVDLSsNAFL3xWeur6tLNYBFclaQKuiy6cVnBPqANZTKdtEMnkzBzI5TQz3DjQhG3fo07/8ZJm4W2Hhg4nHMvc+4JEikMu63s7a+sbm1Xdop7+7tHxWjo7bJk414y0Wy1h3A2q4FIq3UKDk3URzGgWSd4LJXe53nrg2IlaPOE24H9GREqFgFK3U60cUx4zKrDkbVKpuzZ2DrBKvIFUo0BxUvrDmKURV8gkNabnuQn6GdUomOSzcj81PKFsQke8Z6miETd+No8I+dWGZIw1vYpJHP190ZGI2OmUWAn84hm2cvF/7xeiuGNnwmVpMgVW3wUpJgTPL7yVBozlBOLaFMC5uVsDHVlKFtqWxL8JZPXiXtes27rNUfrqN26KOEpzCGVyAB9fQgHtoQgsYxPAMr/DmoPivDsfi9E1p9g5gT9wPn8AiGeRag=</latexit> Abstraction via Adaptive Data Analysis Framework of Dwork et al. (2015) query q 1 S analyst mechanism data set S ∼ P n q i : supp( P ) → [0 , 1] d - queries posed by analyst P - population distribution - sample size n