{ "name": "Machine Learning Blog | ML@CMU | Carnegie Mellon University \u2013 1 article", "role": "AXWindow", "description": "None", "role_description": "standard window", "value": "None", "children": [ { "name": "None", "role": "AXSplitGroup", "description": "None", "role_description": "split group", "value": "None", "children": [ { "name": "None", "role": "AXScrollArea", "description": "None", "role_description": "scroll area", "value": "None", "children": [ { "name": "None", "role": "AXOutline", "description": "None", "role_description": "outline", "value": "None", "children": [ { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Smart Filters", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Smart Filters", "children": [], "bbox": [ 28, 108, 372, 28 ], "visible_bbox": [ 28, 108, 372, 28 ] } ], "bbox": [ 20, 104, 436, 38 ], "visible_bbox": [ 20, 104, 436, 38 ] } ], "bbox": [ 0, 104, 476, 38 ], "visible_bbox": [ 0, 104, 476, 38 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Latest News", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "today selection", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 150, 40, 40 ], "visible_bbox": [ 32, 150, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Latest News", "children": [], "bbox": [ 76, 154, 320, 32 ], "visible_bbox": [ 76, 154, 320, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "1", "children": [], "bbox": [ 396, 154, 52, 30 ], "visible_bbox": [ 396, 154, 52, 30 ] } ], "bbox": [ 20, 142, 436, 56 ], "visible_bbox": [ 20, 142, 436, 56 ] } ], "bbox": [ 0, 142, 476, 56 ], "visible_bbox": [ 0, 142, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Unread Items", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "unread selection", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 206, 40, 40 ], "visible_bbox": [ 32, 206, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Unread Items", "children": [], "bbox": [ 76, 210, 320, 32 ], "visible_bbox": [ 76, 210, 320, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "99", "children": [], "bbox": [ 396, 210, 52, 30 ], "visible_bbox": [ 396, 210, 52, 30 ] } ], "bbox": [ 20, 198, 436, 56 ], "visible_bbox": [ 20, 198, 436, 56 ] } ], "bbox": [ 0, 198, 476, 56 ], "visible_bbox": [ 0, 198, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "None", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "favorite selection", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 262, 40, 40 ], "visible_bbox": [ 32, 262, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Favorites", "children": [], "bbox": [ 76, 266, 368, 32 ], "visible_bbox": [ 76, 266, 368, 32 ] } ], "bbox": [ 20, 254, 436, 56 ], "visible_bbox": [ 20, 254, 436, 56 ] } ], "bbox": [ 0, 254, 476, 56 ], "visible_bbox": [ 0, 254, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Subscriptions", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Subscriptions", "children": [], "bbox": [ 28, 340, 372, 28 ], "visible_bbox": [ 28, 340, 372, 28 ] } ], "bbox": [ 20, 336, 436, 38 ], "visible_bbox": [ 20, 336, 436, 38 ] } ], "bbox": [ 0, 336, 476, 38 ], "visible_bbox": [ 0, 336, 476, 38 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "\"machine learning\" - Google News", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "A3AAC1958952CCE213E3A99604258E", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 382, 40, 40 ], "visible_bbox": [ 32, 382, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "\"machine learning\" - Google News", "children": [], "bbox": [ 76, 386, 320, 32 ], "visible_bbox": [ 76, 386, 320, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "27", "children": [], "bbox": [ 396, 386, 52, 30 ], "visible_bbox": [ 396, 386, 52, 30 ] } ], "bbox": [ 20, 374, 436, 56 ], "visible_bbox": [ 20, 374, 436, 56 ] } ], "bbox": [ 0, 374, 476, 56 ], "visible_bbox": [ 0, 374, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Blog", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "6BECB11E9ADBE43B0A15F67C68E81C", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 438, 40, 40 ], "visible_bbox": [ 32, 438, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Blog", "children": [], "bbox": [ 76, 442, 368, 32 ], "visible_bbox": [ 76, 442, 368, 32 ] } ], "bbox": [ 20, 430, 436, 56 ], "visible_bbox": [ 20, 430, 436, 56 ] } ], "bbox": [ 0, 430, 476, 56 ], "visible_bbox": [ 0, 430, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Machine Learning", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "06481ED9438C8A9F427F01198E7977", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 494, 40, 40 ], "visible_bbox": [ 32, 494, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Machine Learning", "children": [], "bbox": [ 76, 498, 320, 32 ], "visible_bbox": [ 76, 498, 320, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "4", "children": [], "bbox": [ 396, 498, 52, 30 ], "visible_bbox": [ 396, 498, 52, 30 ] } ], "bbox": [ 20, 486, 436, 56 ], "visible_bbox": [ 20, 486, 436, 56 ] } ], "bbox": [ 0, 486, 476, 56 ], "visible_bbox": [ 0, 486, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Machine Learning", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "1A2A0995BC4E5BEFC0597E04C0BB66", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 550, 40, 40 ], "visible_bbox": [ 32, 550, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Machine Learning", "children": [], "bbox": [ 76, 554, 368, 32 ], "visible_bbox": [ 76, 554, 368, 32 ] } ], "bbox": [ 20, 542, 436, 56 ], "visible_bbox": [ 20, 542, 436, 56 ] } ], "bbox": [ 0, 542, 476, 56 ], "visible_bbox": [ 0, 542, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Machine Learning (Theory)", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "0D7F7592C8886465D045C93B3E9FD4", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 606, 40, 40 ], "visible_bbox": [ 32, 606, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Machine Learning (Theory)", "children": [], "bbox": [ 76, 610, 320, 32 ], "visible_bbox": [ 76, 610, 320, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "9", "children": [], "bbox": [ 396, 610, 52, 30 ], "visible_bbox": [ 396, 610, 52, 30 ] } ], "bbox": [ 20, 598, 436, 56 ], "visible_bbox": [ 20, 598, 436, 56 ] } ], "bbox": [ 0, 598, 476, 56 ], "visible_bbox": [ 0, 598, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Machine learning : nature.com subject feeds", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "3A08FAF071AE3B08A7858464A494A9", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 662, 40, 40 ], "visible_bbox": [ 32, 662, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Machine learning : nature.com subject feeds", "children": [], "bbox": [ 76, 666, 320, 32 ], "visible_bbox": [ 76, 666, 320, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "24", "children": [], "bbox": [ 396, 666, 52, 30 ], "visible_bbox": [ 396, 666, 52, 30 ] } ], "bbox": [ 20, 654, 436, 56 ], "visible_bbox": [ 20, 654, 436, 56 ] } ], "bbox": [ 0, 654, 476, 56 ], "visible_bbox": [ 0, 654, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Machine Learning Blog", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "06B425838EB0C9AFE51860D6C69C5A", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 718, 40, 40 ], "visible_bbox": [ 32, 718, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Machine Learning Blog", "children": [], "bbox": [ 76, 722, 368, 32 ], "visible_bbox": [ 76, 722, 368, 32 ] } ], "bbox": [ 20, 710, 436, 56 ], "visible_bbox": [ 20, 710, 436, 56 ] } ], "bbox": [ 0, 710, 476, 56 ], "visible_bbox": [ 0, 710, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Machine Learning Blog | ML@CMU | Carnegie Mellon University", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "4685CDF8E55440EB938CF416F91C0B", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 774, 40, 40 ], "visible_bbox": [ 32, 774, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Machine Learning Blog | ML@CMU | Carnegie Mellon University", "children": [], "bbox": [ 76, 778, 368, 32 ], "visible_bbox": [ 76, 778, 368, 32 ] } ], "bbox": [ 20, 766, 436, 56 ], "visible_bbox": [ 20, 766, 436, 56 ] } ], "bbox": [ 0, 766, 476, 56 ], "visible_bbox": [ 0, 766, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "Machine Learning Times", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "23B110929664670359A372CF7001AE", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 830, 40, 40 ], "visible_bbox": [ 32, 830, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Machine Learning Times", "children": [], "bbox": [ 76, 834, 320, 32 ], "visible_bbox": [ 76, 834, 320, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "6", "children": [], "bbox": [ 396, 834, 52, 30 ], "visible_bbox": [ 396, 834, 52, 30 ] } ], "bbox": [ 20, 822, 436, 56 ], "visible_bbox": [ 20, 822, 436, 56 ] } ], "bbox": [ 0, 822, 476, 56 ], "visible_bbox": [ 0, 822, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "machine learning | TechCrunch", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "0AD22675FA9B5A9063ABC5E1D8959C", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 886, 40, 40 ], "visible_bbox": [ 32, 886, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "machine learning | TechCrunch", "children": [], "bbox": [ 76, 890, 320, 32 ], "visible_bbox": [ 76, 890, 320, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "20", "children": [], "bbox": [ 396, 890, 52, 30 ], "visible_bbox": [ 396, 890, 52, 30 ] } ], "bbox": [ 20, 878, 436, 56 ], "visible_bbox": [ 20, 878, 436, 56 ] } ], "bbox": [ 0, 878, 476, 56 ], "visible_bbox": [ 0, 878, 476, 56 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "outline row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "MachineLearningMastery.com", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXImage", "description": "EA6A09D80696442FCE237C20DFFAC3", "role_description": "image", "value": "None", "children": [], "bbox": [ 32, 942, 40, 40 ], "visible_bbox": [ 32, 942, 40, 40 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "MachineLearningMastery.com", "children": [], "bbox": [ 76, 946, 320, 32 ], "visible_bbox": [ 76, 946, 320, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "9", "children": [], "bbox": [ 396, 946, 52, 30 ], "visible_bbox": [ 396, 946, 52, 30 ] } ], "bbox": [ 20, 934, 436, 56 ], "visible_bbox": [ 20, 934, 436, 56 ] } ], "bbox": [ 0, 934, 476, 56 ], "visible_bbox": [ 0, 934, 476, 56 ] }, { "name": "None", "role": "AXColumn", "description": "None", "role_description": "column", "value": "None", "children": [], "bbox": [ 20, 104, 436, 1210 ], "visible_bbox": [ 20, 104, 436, 1210 ] } ], "bbox": [ 0, 104, 476, 1210 ], "visible_bbox": [ 0, 104, 476, 1210 ] } ], "bbox": [ 0, 104, 476, 1210 ], "visible_bbox": [ 0, 104, 476, 1210 ] }, { "name": "", "role": "AXButton", "description": "None", "role_description": "button", "value": "None", "children": [], "bbox": [ 22, 1342, 32, 32 ], "visible_bbox": [ 22, 1342, 32, 32 ] }, { "name": "", "role": "AXCheckBox", "description": "None", "role_description": "toggle button", "value": "1", "children": [], "bbox": [ 314, 1342, 32, 32 ], "visible_bbox": [ 314, 1342, 32, 32 ] }, { "name": "", "role": "AXCheckBox", "description": "None", "role_description": "toggle button", "value": "0", "children": [], "bbox": [ 368, 1342, 32, 32 ], "visible_bbox": [ 368, 1342, 32, 32 ] }, { "name": "", "role": "AXCheckBox", "description": "None", "role_description": "toggle button", "value": "0", "children": [], "bbox": [ 418, 1342, 32, 32 ], "visible_bbox": [ 418, 1342, 32, 32 ] }, { "name": "", "role": "AXButton", "description": "None", "role_description": "button", "value": "None", "children": [], "bbox": [ 68, 1340, 40, 40 ], "visible_bbox": [ 68, 1340, 40, 40 ] }, { "name": "None", "role": "AXSplitter", "description": "None", "role_description": "splitter", "value": "239.0", "children": [], "bbox": [ 478, 104, 2, 1282 ], "visible_bbox": [ 478, 104, 2, 1282 ] }, { "name": "None", "role": "AXScrollArea", "description": "None", "role_description": "scroll area", "value": "None", "children": [ { "name": "None", "role": "AXTable", "description": "None", "role_description": "table", "value": "None", "children": [ { "name": "None", "role": "AXRow", "description": "None", "role_description": "table row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "None", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "1 month ago", "children": [], "bbox": [ 508, 110, 142, 34 ], "visible_bbox": [ 508, 110, 142, 34 ] } ], "bbox": [ 480, 124, 600, 46 ], "visible_bbox": [ 480, 124, 600, 46 ] } ], "bbox": [ 480, 124, 600, 46 ], "visible_bbox": [ 480, 124, 600, 46 ] }, { "name": "None", "role": "AXRow", "description": "None", "role_description": "table row", "value": "None", "children": [ { "name": "None", "role": "AXCell", "description": "None", "role_description": "cell", "value": "None", "children": [ { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Machine Learning Blog | ML@CMU | Carnegie Mellon University", "children": [], "bbox": [ 662, 378, 184, 28 ], "visible_bbox": [ 662, 378, 184, 28 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "On Noisy Evaluation in Federated Hyperparameter Tuning", "children": [], "bbox": [ 662, 188, 384, 96 ], "visible_bbox": [ 662, 188, 384, 96 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Evaluating models in federated networks is challenging due to factors such as client subsampling, data heterogeneity, and privacy. These factors introduce noise that can affect hyperparameter tuning algorithms and lead to suboptimal model selection. Hyperparameter tuning is critical to the success of cross-device federated learning applications. Unfortunately, federated networks face issues of scale, heterogeneity, and privacy, which introduce noise in the tuning process and make it difficult to faithfully evaluate the performance of various hyperparameters. Our work (MLSys\u201923) explores key sources of noise and surprisingly shows that even small amounts of noise can have a significant impact on tuning methods\u2014reducing the performance of state-of-the-art approaches to that of naive baselines. To address noisy evaluation in such scenarios, we propose a simple and effective approach that leverages public proxy data to boost the evaluation signal. Our work establishes general challenges, baselines, and best practices for future work in federated hyperparameter tuning. Federated Learning: An Overview Cross-device federated learning (FL) is a machine learning setting that considers training a model over a large heterogeneous network of devices such as mobile phones or wearables. Three key factors differentiate FL from traditional centralized learning and distributed learning: Scale. Cross-device refers to FL settings with [\u2026]", "children": [], "bbox": [ 662, 286, 384, 90 ], "visible_bbox": [ 662, 286, 384, 90 ] }, { "name": "None", "role": "AXImage", "description": "", "role_description": "image", "value": "None", "children": [], "bbox": [ 1042, 408, 0, 0 ], "visible_bbox": null }, { "name": "None", "role": "AXImage", "description": "", "role_description": "image", "value": "None", "children": [], "bbox": [ 518, 188, 128, 128 ], "visible_bbox": [ 518, 188, 128, 128 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "29/12/2023, 18:04", "children": [], "bbox": [ 846, 378, 200, 28 ], "visible_bbox": [ 846, 378, 200, 28 ] }, { "name": "None", "role": "AXImage", "description": "", "role_description": "image", "value": "None", "children": [], "bbox": [ 1042, 408, 0, 0 ], "visible_bbox": null } ], "bbox": [ 512, 170, 536, 254 ], "visible_bbox": [ 512, 170, 536, 254 ] } ], "bbox": [ 480, 170, 600, 254 ], "visible_bbox": [ 480, 170, 600, 254 ] }, { "name": "None", "role": "AXColumn", "description": "None", "role_description": "column", "value": "None", "children": [], "bbox": [ 500, 104, 560, 1282 ], "visible_bbox": [ 500, 104, 560, 1282 ] } ], "bbox": [ 480, 104, 600, 1282 ], "visible_bbox": [ 480, 104, 600, 1282 ] } ], "bbox": [ 480, 104, 600, 1282 ], "visible_bbox": [ 480, 104, 600, 1282 ] }, { "name": "None", "role": "AXSplitter", "description": "None", "role_description": "splitter", "value": "300.0", "children": [], "bbox": [ 1080, 104, 2, 1282 ], "visible_bbox": [ 1080, 104, 2, 1282 ] }, { "name": "None", "role": "AXGroup", "description": "None", "role_description": "group", "value": "None", "children": [ { "name": "None", "role": "AXGroup", "description": "None", "role_description": "group", "value": "None", "children": [ { "name": "", "role": "AXScrollArea", "description": "", "role_description": "scroll area", "value": "", "children": [ { "name": "", "role": "AXWebArea", "description": "", "role_description": "HTML content", "value": "", "children": [ { "name": "Machine Learning Blog | ML@CMU | Carnegie Mellon University \u00ad\u2013 Kevin Kuo On Noisy Evaluation in Federated Hyperparameter Tuning Friday, 29 December 2023, 18:04", "role": "AXLink", "description": "", "role_description": "link", "value": "", "children": [ { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Machine Learning Blog | ML@CMU | Carnegie Mellon University \u00ad\u2013 Kevin Kuo", "children": [], "bbox": [ 1120, 148, 986, 36 ], "visible_bbox": [ 1120, 148, 986, 36 ] } ], "bbox": [ 1120, 142, 1168, 50 ], "visible_bbox": [ 1120, 142, 1168, 50 ] }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "On Noisy Evaluation in Federated Hyperparameter Tuning", "children": [], "bbox": [ 1120, 190, 980, 144 ], "visible_bbox": [ 1120, 190, 980, 144 ] } ], "bbox": [ 1120, 190, 1168, 144 ], "visible_bbox": [ 1120, 190, 1168, 144 ] }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Friday, 29 December 2023, 18:04", "children": [], "bbox": [ 1120, 354, 438, 36 ], "visible_bbox": [ 1120, 354, 438, 36 ] } ], "bbox": [ 1120, 348, 1168, 50 ], "visible_bbox": [ 1120, 348, 1168, 50 ] } ], "bbox": [ 1120, 142, 1168, 256 ], "visible_bbox": [ 1120, 142, 1168, 256 ] }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXImage", "description": "", "role_description": "image", "value": "", "children": [], "bbox": [ 1120, 396, 1168, 696 ], "visible_bbox": [ 1120, 396, 1168, 696 ] } ], "bbox": [ 1120, 396, 1168, 718 ], "visible_bbox": [ 1120, 396, 1168, 718 ] }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Evaluating models in federated networks is challenging due to factors such as client subsampling, data heterogeneity, and privacy. These factors introduce noise that can affect hyperparameter tuning algorithms and lead to suboptimal model selection.", "children": [], "bbox": [ 1120, 1150, 1126, 182 ], "visible_bbox": [ 1120, 1150, 1126, 182 ] } ], "bbox": [ 1120, 1144, 1168, 194 ], "visible_bbox": [ 1120, 1144, 1168, 194 ] }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Hyperparameter tuning is critical to the success of cross-device federated learning applications. Unfortunately, federated networks face issues of scale, heterogeneity, and privacy, which introduce noise in the tuning process and make it difficult to faithfully evaluate the performance of various hyperparameters. ", "children": [], "bbox": [ 1120, 1374, 1130, 230 ], "visible_bbox": [ 1120, 1374, 1130, 12 ] }, { "name": "Our work (MLSys\u201923)", "role": "AXLink", "description": "", "role_description": "link", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Our work (MLSys\u201923)", "children": [], "bbox": [ 1192, 1566, 304, 38 ], "visible_bbox": null } ], "bbox": [ 1192, 1566, 304, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " explores key sources of noise and surprisingly shows that even small amounts of noise can have a significant impact on tuning methods\u2014reducing the performance of state-of-the-art approaches to that of naive baselines. To address noisy evaluation in such scenarios, we propose a simple and effective approach that leverages public proxy data to boost the evaluation signal. Our work establishes general challenges, baselines, and best practices for future work in federated hyperparameter tuning.", "children": [], "bbox": [ 1120, 1566, 1166, 326 ], "visible_bbox": null } ], "bbox": [ 1120, 1368, 1168, 530 ], "visible_bbox": [ 1120, 1368, 1168, 18 ] }, { "name": "Federated Learning: An Overview", "role": "AXHeading", "description": "", "role_description": "heading", "value": "2", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Federated Learning: An Overview", "children": [], "bbox": [ 1120, 1942, 560, 44 ], "visible_bbox": null } ], "bbox": [ 1120, 1940, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "In federated learning (FL), user data remains on the device and only model updates are communicated. (Source: Wikipedia)", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXImage", "description": "undefined", "role_description": "image", "value": "", "children": [], "bbox": [ 1120, 2020, 1168, 814 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "In federated learning (FL), user data remains on the device and only model updates are communicated. (Source: Wikipedia)", "children": [], "bbox": [ 1120, 2860, 1160, 70 ], "visible_bbox": null } ], "bbox": [ 1120, 2020, 1168, 914 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "C", "children": [], "bbox": [ 1120, 2970, 24, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "ross-device federated learning", "children": [], "bbox": [ 1142, 2970, 444, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " (FL) is a machine learning setting that considers training a model over a ", "children": [], "bbox": [ 1120, 2970, 1160, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "large heterogeneous network", "children": [], "bbox": [ 1452, 3018, 426, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " of devices such as mobile phones or wearables. Three key factors differentiate FL from traditional centralized learning and distributed learning:", "children": [], "bbox": [ 1120, 3018, 1136, 134 ], "visible_bbox": null } ], "bbox": [ 1120, 2964, 1168, 194 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Scale", "children": [], "bbox": [ 1120, 3194, 86, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ". ", "children": [], "bbox": [ 1204, 3194, 18, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Cross-device", "children": [], "bbox": [ 1220, 3194, 196, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " refers to FL settings with ", "children": [], "bbox": [ 1414, 3194, 370, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "many clients", "children": [], "bbox": [ 1782, 3194, 182, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " with potentially", "children": [], "bbox": [ 1962, 3194, 226, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "limited local resources", "children": [], "bbox": [ 1120, 3242, 324, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " e.g. training a language model across hundreds to millions of mobile phones. These devices have various resource constraints, such as limited upload speed, number of local examples, or computational capability.", "children": [], "bbox": [ 1120, 3242, 1164, 134 ], "visible_bbox": null } ], "bbox": [ 1120, 3188, 1168, 194 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Heterogeneity.", "children": [], "bbox": [ 1120, 3418, 232, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " Traditional distributed ML assumes each worker/client has a random (identically distributed) sample of the training data. In contrast, in FL client datasets may be non-identically distributed, with each user\u2019s data being generated by a distinct underlying distribution.", "children": [], "bbox": [ 1120, 3418, 1118, 182 ], "visible_bbox": null } ], "bbox": [ 1120, 3412, 1168, 194 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Privacy. ", "children": [], "bbox": [ 1120, 3642, 132, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FL offers a baseline level of privacy since raw user data remains local on each client. However, FL is still vulnerable to post-hoc attacks where the public output of the FL algorithm (e.g. a model or its hyperparameters) can be reverse-engineered and leak private user information. A common approach to mitigate such vulnerabilities is to use ", "children": [], "bbox": [ 1120, 3642, 1160, 230 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "differential privacy", "children": [], "bbox": [ 1530, 3834, 266, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ", which aims to mask the contribution of each client. However, differential privacy introduces noise in the aggregate evaluation signal, which can make it difficult to effectively select models.", "children": [], "bbox": [ 1120, 3834, 1134, 182 ], "visible_bbox": null } ], "bbox": [ 1120, 3636, 1168, 386 ], "visible_bbox": null }, { "name": "Federated Hyperparameter Tuning", "role": "AXHeading", "description": "", "role_description": "heading", "value": "3", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Federated Hyperparameter Tuning", "children": [], "bbox": [ 1120, 4068, 548, 42 ], "visible_bbox": null } ], "bbox": [ 1120, 4064, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Appropriately selecting ", "children": [], "bbox": [ 1120, 4128, 342, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "hyperparameters ", "children": [], "bbox": [ 1460, 4128, 256, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "(HPs) is critical to training quality models in FL. Hyperparameters are ", "children": [], "bbox": [ 1120, 4128, 1064, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "user-specified parameters ", "children": [], "bbox": [ 1630, 4176, 392, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "that dictate the process of model training such as the learning rate, local batch size, and number of clients sampled at each round. The problem of tuning HPs is general to machine learning (not just FL). Given an HP search space and search budget, HP tuning methods aim to find a configuration in the search space that optimizes some measure of quality within a constrained budget.", "children": [], "bbox": [ 1120, 4176, 1154, 278 ], "visible_bbox": null } ], "bbox": [ 1120, 4122, 1168, 338 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Let\u2019s first look at an end-to-end FL pipeline that considers both the processes of training and hyperparameter tuning. In cross-device FL, we split the clients into two pools for training and validation. Given a hyperparameter configuration \\((\\lambda_s, \\lambda_c)\\), we train a model using the training clients (explained in section \u201cFL Training\u201d). We then evaluate this model on the validation clients, obtaining an error rate/accuracy metric. We can then use the error rate to adjust the hyperparameters and train a new model.", "children": [], "bbox": [ 1120, 4496, 1168, 326 ], "visible_bbox": null } ], "bbox": [ 1120, 4490, 1168, 338 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "A standard pipeline for tuning hyperparameters in cross-device FL.", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "A standard pipeline for tuning hyperparameters in cross-device FL.", "children": [], "bbox": [ 1120, 5446, 794, 32 ], "visible_bbox": null } ], "bbox": [ 1120, 4858, 1168, 624 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "The diagram above shows two vectors of hyperparameters \\(\\lambda_s, \\lambda_c\\). These correspond to the hyperparameters of two optimizers: one is", "children": [], "bbox": [ 1120, 5518, 1130, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "server-side", "children": [], "bbox": [ 1120, 5614, 166, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " and the other is ", "children": [], "bbox": [ 1284, 5614, 242, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "client-side", "children": [], "bbox": [ 1524, 5614, 154, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ". Next, we describe how these hyperparameters are used during FL training.", "children": [], "bbox": [ 1120, 5614, 1156, 86 ], "visible_bbox": null } ], "bbox": [ 1120, 5512, 1168, 194 ], "visible_bbox": null }, { "name": "FL Training", "role": "AXHeading", "description": "", "role_description": "heading", "value": "2", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FL Training", "children": [], "bbox": [ 1120, 5750, 186, 44 ], "visible_bbox": null } ], "bbox": [ 1120, 5748, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "A typical FL algorithm consists of several ", "children": [], "bbox": [ 1120, 5812, 592, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "rounds", "children": [], "bbox": [ 1710, 5812, 104, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " of training where each client performs local training followed by aggregation of the client updates. In our work, we experiment with a general framework called ", "children": [], "bbox": [ 1120, 5812, 1162, 134 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FedOPT", "children": [], "bbox": [ 1806, 5912, 106, 34 ], "visible_bbox": null } ], "bbox": [ 1800, 5906, 118, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " which was presented in", "children": [], "bbox": [ 1916, 5908, 346, 38 ], "visible_bbox": null }, { "name": "Adaptive Federated Optimization (Reddi et al. 2021)", "role": "AXLink", "description": "", "role_description": "link", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Adaptive Federated Optimization (Reddi et al. 2021)", "children": [], "bbox": [ 1120, 5958, 740, 38 ], "visible_bbox": null } ], "bbox": [ 1120, 5958, 740, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ". We outline the per-round procedure of ", "children": [], "bbox": [ 1120, 5958, 1116, 86 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FedOPT", "children": [], "bbox": [ 1318, 6010, 106, 34 ], "visible_bbox": null } ], "bbox": [ 1314, 6004, 118, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " below:", "children": [], "bbox": [ 1428, 6006, 106, 38 ], "visible_bbox": null } ], "bbox": [ 1120, 5806, 1168, 246 ], "visible_bbox": null }, { "name": "", "role": "AXList", "description": "", "role_description": "content list", "value": "", "children": [ { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "1", "children": [], "bbox": [ 1154, 6088, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "The server broadcasts the model \\(\\theta\\) to a sampled subset of \\(K\\) clients.", "children": [], "bbox": [ 1200, 6088, 1010, 86 ], "visible_bbox": null } ], "bbox": [ 1200, 6082, 1088, 98 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "2", "children": [], "bbox": [ 1154, 6184, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Each client (in parallel) trains \\(\\theta\\) on their local data \\(X_k\\) using", "children": [], "bbox": [ 1200, 6184, 996, 38 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "ClientOPT", "children": [], "bbox": [ 1206, 6236, 158, 34 ], "visible_bbox": null } ], "bbox": [ 1200, 6230, 170, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "and obtains an updated model \\(\\theta_k\\).", "children": [], "bbox": [ 1376, 6232, 612, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 6178, 1088, 100 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "3", "children": [], "bbox": [ 1154, 6282, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Each client sends \\(\\theta_k\\) back to the server.", "children": [], "bbox": [ 1200, 6282, 692, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 6276, 1088, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "4", "children": [], "bbox": [ 1154, 6330, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "The server averages all the received models \\\\(\\theta\u2019 = \\frac{1}{K} \\sum_k p_k\\theta_k\\).", "children": [], "bbox": [ 1200, 6330, 1052, 86 ], "visible_bbox": null } ], "bbox": [ 1200, 6324, 1088, 98 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "5", "children": [], "bbox": [ 1154, 6426, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "To update \\(\\theta\\), the server computes the difference \\(\\theta \u2013 \\theta\u2019\\) and feeds it as a pseudo-gradient into ", "children": [], "bbox": [ 1200, 6426, 1014, 86 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "ServerOPT", "children": [], "bbox": [ 1932, 6478, 158, 34 ], "visible_bbox": null } ], "bbox": [ 1926, 6472, 170, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " (rather than computing a gradient w.r.t. some loss function).", "children": [], "bbox": [ 1200, 6474, 1012, 88 ], "visible_bbox": null } ], "bbox": [ 1200, 6420, 1088, 148 ], "visible_bbox": null } ], "bbox": [ 1120, 6082, 1168, 486 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "The FedOPT framework and the five hyperparameters (\\(\\lambda_s, \\lambda_c\\)) we consider tuning. (Source: edited from Wikipedia)", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "The FedOPT framework and the five hyperparameters (\\(\\lambda_s, \\lambda_c\\)) we consider tuning. (Source: edited from Wikipedia)", "children": [], "bbox": [ 1120, 7096, 1108, 70 ], "visible_bbox": null } ], "bbox": [ 1120, 6598, 1168, 572 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Steps 2 and 5 of ", "children": [], "bbox": [ 1120, 7206, 244, 38 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FedOPT", "children": [], "bbox": [ 1368, 7210, 106, 34 ], "visible_bbox": null } ], "bbox": [ 1362, 7204, 118, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " each require a gradient-based", "children": [], "bbox": [ 1478, 7206, 444, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " optimization algorithm", "children": [], "bbox": [ 1920, 7206, 332, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "(called ", "children": [], "bbox": [ 1120, 7256, 106, 38 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "ClientOPT", "children": [], "bbox": [ 1230, 7260, 158, 34 ], "visible_bbox": null } ], "bbox": [ 1224, 7254, 170, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " and ", "children": [], "bbox": [ 1392, 7256, 72, 38 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "ServerOPT", "children": [], "bbox": [ 1468, 7260, 158, 34 ], "visible_bbox": null } ], "bbox": [ 1462, 7254, 170, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ") which specify how to update \\(\\theta\\) given some update vector. In our work, we focus on an instantiation of ", "children": [], "bbox": [ 1120, 7256, 1158, 88 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FedOPT", "children": [], "bbox": [ 2046, 7310, 106, 34 ], "visible_bbox": null } ], "bbox": [ 2040, 7304, 118, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " called", "children": [], "bbox": [ 2156, 7306, 96, 38 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FedAdam", "children": [], "bbox": [ 1126, 7360, 122, 34 ], "visible_bbox": null } ], "bbox": [ 1120, 7354, 134, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ", which uses ", "children": [], "bbox": [ 1252, 7356, 188, 38 ], "visible_bbox": null }, { "name": "Adam (Kingma and Ba 2014)", "role": "AXLink", "description": "", "role_description": "link", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Adam (Kingma and Ba 2014)", "children": [], "bbox": [ 1438, 7356, 410, 38 ], "visible_bbox": null } ], "bbox": [ 1438, 7356, 410, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " as ", "children": [], "bbox": [ 1846, 7356, 50, 38 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "ServerOPT", "children": [], "bbox": [ 1900, 7360, 158, 34 ], "visible_bbox": null } ], "bbox": [ 1894, 7354, 170, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " and SGD as", "children": [], "bbox": [ 2062, 7356, 178, 38 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "ClientOPT", "children": [], "bbox": [ 1126, 7410, 158, 34 ], "visible_bbox": null } ], "bbox": [ 1120, 7404, 170, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ". We focus on tuning five ", "children": [], "bbox": [ 1288, 7406, 362, 38 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FedAdam", "children": [], "bbox": [ 1654, 7410, 124, 34 ], "visible_bbox": null } ], "bbox": [ 1648, 7404, 136, 46 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " hyperparameters: two for client training (SGD\u2019s learning rate and batch size) and three for server aggregation (Adam\u2019s learning rate, 1st-moment decay, and 2nd-moment decay).", "children": [], "bbox": [ 1120, 7406, 1120, 136 ], "visible_bbox": null } ], "bbox": [ 1120, 7200, 1168, 348 ], "visible_bbox": null }, { "name": "FL Evaluation", "role": "AXHeading", "description": "", "role_description": "heading", "value": "2", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FL Evaluation", "children": [], "bbox": [ 1120, 7592, 224, 44 ], "visible_bbox": null } ], "bbox": [ 1120, 7590, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Now, we discuss how ", "children": [], "bbox": [ 1120, 7654, 314, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FL settings introduce noise to model evaluation", "children": [], "bbox": [ 1432, 7654, 680, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ". Consider the following example below. We have \\(K=4\\) configurations (grey, blue, red, green) and we want to figure out which configuration has the best average accuracy across \\(N=5\\) clients. More specifically, each \u201cconfiguration\u201d is a set of HP values (learning rate, batch size, etc.) that are fed into an ", "children": [], "bbox": [ 1120, 7654, 1158, 230 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "FL training algorithm", "children": [], "bbox": [ 1990, 7846, 298, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "(more details in the next section). This produces a model we can evaluate. If we can evaluate every model on every client then our evaluation is ", "children": [], "bbox": [ 1120, 7894, 1136, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "noiseless", "children": [], "bbox": [ 2022, 7942, 136, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ". In this case, we would be able to accurately determine that the green model performs the best. However, generating all the evaluations as shown below is not practical, as evaluation costs scale with both the number of configurations and clients.", "children": [], "bbox": [ 1120, 7942, 1158, 182 ], "visible_bbox": null } ], "bbox": [ 1120, 7648, 1168, 482 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "HP tuning without noise. Every configuration is evaluated on every client, which allows us to find the best (green) configuration.", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "HP tuning without noise. Every configuration is evaluated on every client, which allows us to find the best (green) configuration.", "children": [], "bbox": [ 1120, 8856, 1140, 70 ], "visible_bbox": null } ], "bbox": [ 1120, 8160, 1168, 770 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Below, we show an evaluation procedure that is more realistic in FL. As the primary challenge in cross-device FL is ", "children": [], "bbox": [ 1120, 8966, 1060, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "scale", "children": [], "bbox": [ 1686, 9014, 76, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ", we evaluate models using only a random ", "children": [], "bbox": [ 1120, 9014, 1120, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "subsample of clients", "children": [], "bbox": [ 1238, 9062, 322, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ". This is shown in the figure by red \u2018X\u2019s and shaded-out phones. We cover three additional sources of noise in FL which can negatively interact with subsampling and introduce even more noise into the evaluation procedure:", "children": [], "bbox": [ 1120, 9062, 1140, 182 ], "visible_bbox": null } ], "bbox": [ 1120, 8960, 1168, 290 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Data heterogeneity.", "children": [], "bbox": [ 1120, 9286, 308, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " FL clients may have non-identically distributed data, meaning that the evaluations on various models can differ between clients. This is shown by the different histograms next to each client. Data heterogeneity is intrinsic to FL and is critical for our observations on noisy evaluation; if all clients had identical datasets, there would be no need to sample more than one client.", "children": [], "bbox": [ 1120, 9286, 1142, 230 ], "visible_bbox": null } ], "bbox": [ 1120, 9280, 1168, 242 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Systems heterogeneity.", "children": [], "bbox": [ 1120, 9558, 370, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " In addition to data heterogeneity, clients may have heterogeneous system capabilities. For example, some clients have better network reception and computational hardware, which allows them to participate in training and evaluation more frequently. This biases performance towards these clients, leading to a poor overall model.", "children": [], "bbox": [ 1120, 9558, 1156, 230 ], "visible_bbox": null } ], "bbox": [ 1120, 9552, 1168, 242 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Differential privacy.", "children": [], "bbox": [ 1120, 9830, 306, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " Using the evaluation output (i.e. the top-performing model), a malicious party can infer whether or not a particular client participated in the FL procedure. At a high level, differential privacy aims to mask user contributions by adding noise to the aggregate evaluation metric. However, this additional noise can make it difficult to faithfully evaluate HP configurations.", "children": [], "bbox": [ 1120, 9830, 1150, 230 ], "visible_bbox": null } ], "bbox": [ 1120, 9824, 1168, 242 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "figure", "value": "", "children": [], "bbox": [ 1120, 10096, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "In the figure above, evaluations can lead to suboptimal model selection when we consider client subsampling, data heterogeneity, and differential privacy. The combination of all these factors leads us to incorrectly choose the red model over the green one.", "children": [], "bbox": [ 1120, 10182, 1168, 182 ], "visible_bbox": null } ], "bbox": [ 1120, 10176, 1168, 194 ], "visible_bbox": null }, { "name": "Experimental Results", "role": "AXHeading", "description": "", "role_description": "heading", "value": "2", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Experimental Results", "children": [], "bbox": [ 1120, 10414, 356, 44 ], "visible_bbox": null } ], "bbox": [ 1120, 10412, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "The first goal of our work is to investigate the impact of four sources of noisy evaluation that we outlined in the section \u201cFL Evaluation\u201d. In more detail, these are our research questions:", "children": [], "bbox": [ 1120, 10476, 1166, 134 ], "visible_bbox": null } ], "bbox": [ 1120, 10470, 1168, 146 ], "visible_bbox": null }, { "name": "", "role": "AXList", "description": "", "role_description": "content list", "value": "", "children": [ { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "1", "children": [], "bbox": [ 1154, 10652, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "How does ", "children": [], "bbox": [ 1200, 10652, 152, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "subsampling ", "children": [], "bbox": [ 1350, 10652, 210, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "validation clients affect HP tuning performance?", "children": [], "bbox": [ 1558, 10652, 688, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 10646, 1088, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "2", "children": [], "bbox": [ 1154, 10700, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "How do the following factors interact with/exacerbate issues of subsampling?", "children": [], "bbox": [ 1200, 10700, 1036, 86 ], "visible_bbox": null }, { "name": "", "role": "AXList", "description": "", "role_description": "content list", "value": "", "children": [ { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u25e6", "children": [], "bbox": [ 1244, 10796, 16, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "data heterogeneity", "children": [], "bbox": [ 1280, 10796, 296, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " (shuffling validation clients\u2019 datasets)", "children": [], "bbox": [ 1574, 10796, 544, 38 ], "visible_bbox": null } ], "bbox": [ 1280, 10790, 1008, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u25e6", "children": [], "bbox": [ 1244, 10844, 16, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "systems heterogeneity", "children": [], "bbox": [ 1280, 10844, 356, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " (biased client subsampling)", "children": [], "bbox": [ 1634, 10844, 406, 38 ], "visible_bbox": null } ], "bbox": [ 1280, 10838, 1008, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u25e6", "children": [], "bbox": [ 1244, 10892, 16, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "privacy ", "children": [], "bbox": [ 1280, 10892, 124, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "(adding Laplace noise to the aggregate evaluation)", "children": [], "bbox": [ 1402, 10892, 728, 38 ], "visible_bbox": null } ], "bbox": [ 1280, 10886, 1008, 50 ], "visible_bbox": null } ], "bbox": [ 1200, 10790, 1088, 146 ], "visible_bbox": null } ], "bbox": [ 1200, 10694, 1088, 242 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "3", "children": [], "bbox": [ 1154, 10940, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "In noisy settings, how do SOTA methods compare to simple baselines?", "children": [], "bbox": [ 1200, 10940, 1010, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 10934, 1088, 50 ], "visible_bbox": null } ], "bbox": [ 1120, 10646, 1168, 338 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Surprisingly, we show that state-of-the-art HP tuning methods can perform catastrophically poorly, even worse than simple baselines (e.g., random search). While we only show results for CIFAR10, results on three other datasets (FEMNIST, StackOverflow, and Reddit) can be found in our paper. CIFAR10 is partitioned such that each client has at most two out of the ten total labels.", "children": [], "bbox": [ 1120, 11020, 1144, 230 ], "visible_bbox": null } ], "bbox": [ 1120, 11014, 1168, 242 ], "visible_bbox": null }, { "name": "Noise hurts random search", "role": "AXHeading", "description": "", "role_description": "heading", "value": "3", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Noise hurts random search", "children": [], "bbox": [ 1120, 11302, 426, 42 ], "visible_bbox": null } ], "bbox": [ 1120, 11298, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "This section investigates questions 1 and 2 using ", "children": [], "bbox": [ 1120, 11362, 708, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "random search (RS)", "children": [], "bbox": [ 1826, 11362, 310, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " as the hyperparameter tuning method. RS is a simple baseline that ", "children": [], "bbox": [ 1120, 11362, 1112, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "randomly samples", "children": [], "bbox": [ 1982, 11410, 264, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "several HP configurations, trains a model for each one, and returns the highest-performing model (i.e. the example in \u201cFL Evaluation\u201d, if the configurations were sampled independently from the same distribution). Generally, each hyperparameter value is sampled from a (log) uniform or normal distribution.", "children": [], "bbox": [ 1120, 11458, 1140, 182 ], "visible_bbox": null } ], "bbox": [ 1120, 11356, 1168, 290 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "Random search with varying only client subsampling (left) and varying both client subsampling and data heterogeneity (right).", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Random search with varying only client subsampling (left) and varying both client subsampling and data heterogeneity (right).", "children": [], "bbox": [ 1120, 11728, 1122, 70 ], "visible_bbox": null } ], "bbox": [ 1120, 11676, 1168, 126 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Client subsampling. ", "children": [], "bbox": [ 1120, 11838, 318, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "We run RS while varying the client subsampling rate from a single client to the full validation client pool. \u201cBest HPs\u201d indicates the best HPs found across all trials of RS. ", "children": [], "bbox": [ 1120, 11838, 1156, 134 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "As we subsample less clients (left), random search performs worse (higher error rate).", "children": [], "bbox": [ 1120, 11934, 1080, 86 ], "visible_bbox": null } ], "bbox": [ 1120, 11832, 1168, 194 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Data heterogeneity. ", "children": [], "bbox": [ 1120, 12062, 314, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "We run RS on three separate validation partitions with varying degrees of data heterogeneity based on the label distributions on each client. ", "children": [], "bbox": [ 1120, 12062, 1128, 134 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Client subsampling generally harms performance but has a greater impact on performance when the data is heterogeneous (IID Fraction = 0 vs. 1).", "children": [], "bbox": [ 1120, 12158, 1130, 134 ], "visible_bbox": null } ], "bbox": [ 1120, 12056, 1168, 242 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "Random search with varying systems heterogeneity (left) and privacy budget (right). Both factors interact negatively with client subsampling.", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Random search with varying systems heterogeneity (left) and privacy budget (right). Both factors interact negatively with client subsampling.", "children": [], "bbox": [ 1120, 12380, 1154, 70 ], "visible_bbox": null } ], "bbox": [ 1120, 12328, 1168, 126 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Systems heterogeneity.", "children": [], "bbox": [ 1120, 12490, 370, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " We run RS and bias the client sampling to reflect four degrees of systems heterogeneity. Based on the model that is currently being evaluated, we assign a higher probability of sampling clients who perform well on this model. ", "children": [], "bbox": [ 1120, 12490, 1158, 182 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Sampling bias leads to worse performance since the biased evaluations are overly optimistic and do not reflect performance over the entire validation pool.", "children": [], "bbox": [ 1120, 12634, 1126, 134 ], "visible_bbox": null } ], "bbox": [ 1120, 12484, 1168, 290 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Privacy.", "children": [], "bbox": [ 1120, 12810, 124, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " We run RS with 5 different evaluation ", "children": [], "bbox": [ 1242, 12810, 546, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "privacy budgets", "children": [], "bbox": [ 1786, 12810, 254, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " \\(\\varepsilon\\). We add noise sampled from \\(\\text{Lap}(M/(\\varepsilon |S|))\\) to the aggregate evaluation, where \\(M\\) is the number of evaluations (16), \\(\\varepsilon\\) is the privacy budget (each curve), and \\(|S|\\) is the number of clients sampled for an evaluation (x-axis). ", "children": [], "bbox": [ 1120, 12810, 1130, 230 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "A smaller privacy budget requires sampling a larger raw number of clients to achieve reasonable performance.", "children": [], "bbox": [ 1120, 13002, 1142, 86 ], "visible_bbox": null } ], "bbox": [ 1120, 12804, 1168, 290 ], "visible_bbox": null }, { "name": "Noise hurts complex methods more than RS", "role": "AXHeading", "description": "", "role_description": "heading", "value": "3", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Noise hurts complex methods more than RS", "children": [], "bbox": [ 1120, 13140, 692, 42 ], "visible_bbox": null } ], "bbox": [ 1120, 13136, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Seeing that noise adversely affects random search, we now focus on question 3: Do the same observations hold for more complex tuning methods? In the next experiment, we compare 4 representative HP tuning methods.", "children": [], "bbox": [ 1120, 13200, 1152, 134 ], "visible_bbox": null } ], "bbox": [ 1120, 13194, 1168, 146 ], "visible_bbox": null }, { "name": "", "role": "AXList", "description": "", "role_description": "content list", "value": "", "children": [ { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u2022", "children": [], "bbox": [ 1164, 13376, 16, 38 ], "visible_bbox": null }, { "name": "Random Search (RS)", "role": "AXLink", "description": "", "role_description": "link", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Random Search (RS)", "children": [], "bbox": [ 1200, 13376, 298, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 13376, 300, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " is a naive baseline.", "children": [], "bbox": [ 1496, 13376, 278, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 13370, 1088, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u2022", "children": [], "bbox": [ 1164, 13424, 16, 38 ], "visible_bbox": null }, { "name": "Tree-Structured Parzen Estimator (TPE)", "role": "AXLink", "description": "", "role_description": "link", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Tree-Structured Parzen Estimator (TPE)", "children": [], "bbox": [ 1200, 13424, 570, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 13424, 570, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " is a ", "children": [], "bbox": [ 1768, 13424, 66, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "selection-based", "children": [], "bbox": [ 1832, 13424, 236, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " method. These methods build a ", "children": [], "bbox": [ 1200, 13424, 996, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "surrogate model", "children": [], "bbox": [ 1534, 13472, 240, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " that predicts the ", "children": [], "bbox": [ 1772, 13472, 254, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "performance of various hyperparameters", "children": [], "bbox": [ 1200, 13472, 1050, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " rather than predictions for the task at hand (e.g. image or language data).", "children": [], "bbox": [ 1200, 13520, 1054, 86 ], "visible_bbox": null } ], "bbox": [ 1200, 13418, 1088, 194 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u2022", "children": [], "bbox": [ 1164, 13616, 16, 38 ], "visible_bbox": null }, { "name": "Hyperband (HB)", "role": "AXLink", "description": "", "role_description": "link", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Hyperband (HB)", "children": [], "bbox": [ 1200, 13616, 234, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 13616, 234, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " is an ", "children": [], "bbox": [ 1432, 13616, 84, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "allocation-based", "children": [], "bbox": [ 1514, 13616, 244, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " method. These methods ", "children": [], "bbox": [ 1756, 13616, 370, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "allocate more resources to the most promising configurations", "children": [], "bbox": [ 1200, 13616, 1038, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ". Hyperband initially samples a large number of configurations but stops training most of them after the first few rounds.", "children": [], "bbox": [ 1200, 13664, 1054, 134 ], "visible_bbox": null } ], "bbox": [ 1200, 13610, 1088, 194 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u2022", "children": [], "bbox": [ 1164, 13808, 16, 38 ], "visible_bbox": null }, { "name": "Bayesian Optimization + Hyperband (BOHB)", "role": "AXLink", "description": "", "role_description": "link", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Bayesian Optimization + Hyperband (BOHB)", "children": [], "bbox": [ 1200, 13808, 632, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 13808, 632, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " is a combined method that uses both the sampling strategy of TPE and the partial evaluations of HB.", "children": [], "bbox": [ 1200, 13808, 1044, 86 ], "visible_bbox": null } ], "bbox": [ 1200, 13802, 1088, 98 ], "visible_bbox": null } ], "bbox": [ 1120, 13370, 1168, 530 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "Examples of (a) selection-based and (b) allocation-based HP tuning methods. (a) uses a surrogate model of the search space to sample the next configuration (numbered in order of exploration), while (b) randomly samples many configurations and adaptively allocates resources to the most promising ones. (Source: Hyperband (Li et al. 2018) )", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Examples of (a) selection-based and (b) allocation-based HP tuning methods. (a) uses a surrogate model of the search space to sample the next configuration (numbered in order of exploration), while (b) randomly samples many configurations and adaptively allocates resources to the most promising ones. (Source: ", "children": [], "bbox": [ 1120, 13982, 1148, 146 ], "visible_bbox": null }, { "name": "Hyperband (Li et al. 2018)", "role": "AXLink", "description": "", "role_description": "link", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Hyperband (Li et al. 2018)", "children": [], "bbox": [ 1566, 14096, 308, 32 ], "visible_bbox": null } ], "bbox": [ 1566, 14096, 308, 32 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ")", "children": [], "bbox": [ 1872, 14096, 12, 32 ], "visible_bbox": null } ], "bbox": [ 1120, 13930, 1168, 202 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "We report the error rate of each HP tuning method (y-axis) at a given budget of rounds (x-axis). ", "children": [], "bbox": [ 1120, 14168, 1134, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Surprisingly, we find that the relative ranking of these methods can be reversed when the ", "children": [], "bbox": [ 1120, 14216, 1062, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "evaluation is noisy. With noise, the performance of all methods degrades, but the degradation is particularly extreme for HB and BOHB. Intuitively, this is because these two methods already inject noise into the HP tuning procedure via early stopping which interacts poorly with additional sources of noise. Therefore, these results indicate a need for HP tuning methods that are specialized for FL, as many of the guiding principles for traditional hyperparameter tuning may not be effective at handling noisy evaluation in FL.", "children": [], "bbox": [ 1120, 14264, 1154, 374 ], "visible_bbox": null } ], "bbox": [ 1120, 14162, 1168, 482 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "We compare 4 HP tuning methods in noiseless vs. noisy FL settings. In the noiseless setting (left), we always sample all the validation clients and do not consider privacy. In the noisy setting (right), we sample 1% of validation clients and have a generous privacy budget of \\(\\varepsilon=100\\).", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "We compare 4 HP tuning methods in noiseless vs. noisy FL settings. In the noiseless setting (left), we always sample all the validation clients and do not consider privacy. In the noisy setting (right), we sample 1% of validation clients and have a generous privacy budget of \\(\\varepsilon=100\\).", "children": [], "bbox": [ 1120, 14726, 1160, 108 ], "visible_bbox": null } ], "bbox": [ 1120, 14674, 1168, 164 ], "visible_bbox": null }, { "name": "Proxy evaluation outperforms noisy evaluation", "role": "AXHeading", "description": "", "role_description": "heading", "value": "3", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Proxy evaluation outperforms noisy evaluation", "children": [], "bbox": [ 1120, 14884, 730, 42 ], "visible_bbox": null } ], "bbox": [ 1120, 14880, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "In practical FL settings, a practitioner may have access to ", "children": [], "bbox": [ 1120, 14944, 828, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "public proxy data", "children": [], "bbox": [ 1946, 14944, 250, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "which can be used to train models and select hyperparameters. However, given two distinct datasets, it is unclear how well hyperparameters can transfer between them. First, we explore the effectiveness of hyperparameter transfer between four datasets. Below, we see that the CIFAR10-FEMNIST and StackOverflow-Reddit pairs (top left, bottom right) show the clearest transfer between the two datasets. One likely reason for this is that these task pairs use the same model architecture: CIFAR10 and FEMNIST are both image classification tasks while StackOverflow and Reddit are next-word prediction tasks.", "children": [], "bbox": [ 1120, 14992, 1168, 374 ], "visible_bbox": null } ], "bbox": [ 1120, 14938, 1168, 434 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "We experimented with 4 datasets in our work (CIFAR10, FEMNIST, StackOverflow, and Reddit). For each pair of datasets, we randomly sample 128 configurations and plot each configuration at the coordinates corresponding to the error rate on the two datasets.", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "We experimented with 4 datasets in our work (CIFAR10, FEMNIST, StackOverflow, and Reddit). For each pair of datasets, we randomly sample 128 configurations and plot each configuration at the coordinates corresponding to the error rate on the two datasets.", "children": [], "bbox": [ 1120, 15454, 1162, 108 ], "visible_bbox": null } ], "bbox": [ 1120, 15402, 1168, 164 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Given the appropriate proxy dataset, we show that a simple method called ", "children": [], "bbox": [ 1120, 15602, 1068, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "one-shot proxy random search", "children": [], "bbox": [ 1120, 15602, 1138, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " can perform extremely well. The algorithm has two steps:", "children": [], "bbox": [ 1120, 15650, 1108, 86 ], "visible_bbox": null } ], "bbox": [ 1120, 15596, 1168, 146 ], "visible_bbox": null }, { "name": "", "role": "AXList", "description": "", "role_description": "content list", "value": "", "children": [ { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "1", "children": [], "bbox": [ 1154, 15778, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Run a random search using the ", "children": [], "bbox": [ 1200, 15778, 454, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "proxy data", "children": [], "bbox": [ 1652, 15778, 154, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " to both train and evaluate HPs. We assume the proxy data is both ", "children": [], "bbox": [ 1200, 15778, 1054, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "public and server-side", "children": [], "bbox": [ 1692, 15826, 324, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ", so we can always evaluate HPs without subsampling clients or adding privacy noise.", "children": [], "bbox": [ 1200, 15826, 1080, 86 ], "visible_bbox": null } ], "bbox": [ 1200, 15772, 1088, 146 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "2", "children": [], "bbox": [ 1154, 15922, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "The output configuration from 1. is used to train a model on the ", "children": [], "bbox": [ 1200, 15922, 910, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "training client data", "children": [], "bbox": [ 1200, 15922, 1018, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": ". Since we pass only a single configuration to this step, ", "children": [], "bbox": [ 1350, 15970, 790, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "validation client data", "children": [], "bbox": [ 1200, 15970, 1078, 86 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "does not affect ", "children": [], "bbox": [ 1358, 16018, 228, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "hyperparameter selection at all.", "children": [], "bbox": [ 1584, 16018, 454, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 15916, 1088, 146 ], "visible_bbox": null } ], "bbox": [ 1120, 15772, 1168, 290 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "In each experiment, we choose one of these datasets to be partitioned among the clients and use the other three datasets as server-side proxy datasets. Our results show that ", "children": [], "bbox": [ 1120, 16098, 1126, 134 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "proxy data can be an effective solution.", "children": [], "bbox": [ 1372, 16194, 610, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Even if the proxy dataset is not an ideal match for the public data, it may be the only available solution under a strict privacy budget.", "children": [], "bbox": [ 1120, 16194, 1126, 134 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " This is shown in the FEMNIST plot where the orange/red lines (text datasets) perform similarly to the \\(\\varepsilon=10\\) curve.", "children": [], "bbox": [ 1120, 16290, 1166, 134 ], "visible_bbox": null } ], "bbox": [ 1120, 16092, 1168, 338 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "We compare tuning HPs using noisy evaluations on the private dataset (with 1% client subsampling and varying the privacy budget \\(\\varepsilon\\) versus noiseless evaluations on the proxy dataset. The proxy HP tuning methods appear as horizontal lines because they are one-shot.", "role_description": "figure", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "We compare tuning HPs using noisy evaluations on the private dataset (with 1% client subsampling and varying the privacy budget \\(\\varepsilon\\) versus noiseless evaluations on the proxy dataset. The proxy HP tuning methods appear as horizontal lines because they are one-shot.", "children": [], "bbox": [ 1120, 16512, 1126, 108 ], "visible_bbox": null } ], "bbox": [ 1120, 16460, 1168, 164 ], "visible_bbox": null }, { "name": "Conclusion", "role": "AXHeading", "description": "", "role_description": "heading", "value": "2", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Conclusion", "children": [], "bbox": [ 1120, 16668, 188, 44 ], "visible_bbox": null } ], "bbox": [ 1120, 16666, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "In conclusion, our study suggests several best practices for federated HP tuning:", "children": [], "bbox": [ 1120, 16730, 1158, 38 ], "visible_bbox": null } ], "bbox": [ 1120, 16724, 1168, 50 ], "visible_bbox": null }, { "name": "", "role": "AXList", "description": "", "role_description": "content list", "value": "", "children": [ { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u2022", "children": [], "bbox": [ 1164, 16810, 16, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Use simple HP tuning methods.", "children": [], "bbox": [ 1200, 16810, 452, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 16804, 1088, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u2022", "children": [], "bbox": [ 1164, 16858, 16, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Sample a sufficiently large number of validation clients.", "children": [], "bbox": [ 1200, 16858, 790, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 16852, 1088, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u2022", "children": [], "bbox": [ 1164, 16906, 16, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Evaluate a representative set of clients.", "children": [], "bbox": [ 1200, 16906, 562, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 16900, 1088, 50 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "\u2022", "children": [], "bbox": [ 1164, 16954, 16, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "If available, proxy data can be an effective solution.", "children": [], "bbox": [ 1200, 16954, 730, 38 ], "visible_bbox": null } ], "bbox": [ 1200, 16948, 1088, 50 ], "visible_bbox": null } ], "bbox": [ 1120, 16804, 1168, 194 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Furthermore, we identify several directions for future work in federated HP tuning:", "children": [], "bbox": [ 1120, 17034, 1064, 86 ], "visible_bbox": null } ], "bbox": [ 1120, 17028, 1168, 98 ], "visible_bbox": null }, { "name": "", "role": "AXList", "description": "", "role_description": "content list", "value": "", "children": [ { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "1", "children": [], "bbox": [ 1154, 17162, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Tailoring HP tuning methods for differential privacy and FL.", "children": [], "bbox": [ 1200, 17162, 912, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " Early stopping methods are inherently noisy/biased and the large number of evaluations they use is at odds with privacy. Another useful direction is to investigate HP methods specific to noisy evaluation.", "children": [], "bbox": [ 1200, 17162, 1086, 182 ], "visible_bbox": null } ], "bbox": [ 1200, 17156, 1088, 194 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "2", "children": [], "bbox": [ 1154, 17354, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "More detailed cost evaluation.", "children": [], "bbox": [ 1200, 17354, 468, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " In our work, we only considered the number of training rounds as our resource budget. However, practical FL settings consider a wide variety of costs, such as total communication, amount of local training, or total time to train a model.", "children": [], "bbox": [ 1200, 17354, 1040, 182 ], "visible_bbox": null } ], "bbox": [ 1200, 17348, 1088, 194 ], "visible_bbox": null }, { "name": "", "role": "AXGroup", "description": "", "role_description": "group", "value": "", "children": [ { "name": "", "role": "AXListMarker", "description": "", "role_description": "list marker", "value": "3", "children": [], "bbox": [ 1154, 17546, 38, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": "Combining proxy and client data for HP tuning.", "children": [], "bbox": [ 1200, 17546, 724, 38 ], "visible_bbox": null }, { "name": "", "role": "AXStaticText", "description": "", "role_description": "text", "value": " A key issue of using public proxy data for HP tuning is that the best proxy dataset is not known in advance. One direction to address this is to design methods that combine public and private evaluations to mitigate bias from proxy data and noise from private data. Another promising direction is to rely on the abundance of public data and design a method that can select the best proxy dataset.", "children": [], "bbox": [ 1200, 17546, 1082, 326 ], "visible_bbox": null } ], "bbox": [ 1200, 17540, 1088, 338 ], "visible_bbox": null } ], "bbox": [ 1120, 17156, 1168, 722 ], "visible_bbox": null } ], "bbox": [ 1082, 104, 1244, 17842 ], "visible_bbox": [ 1082, 104, 1244, 1282 ] }, { "name": "", "role": "AXScrollBar", "description": "", "role_description": "scroll bar", "value": "0.0", "children": [], "bbox": [ 2294, 208, 32, 1282 ], "visible_bbox": [ 2294, 208, 32, 1178 ] } ], "bbox": [ 1082, 104, 1244, 1282 ], "visible_bbox": [ 1082, 104, 1244, 1282 ] } ], "bbox": [ 1082, 0, 1244, 1386 ], "visible_bbox": [ 1082, 0, 1244, 1386 ] } ], "bbox": [ 1082, 0, 1244, 1386 ], "visible_bbox": [ 1082, 0, 1244, 1386 ] } ], "bbox": [ 0, 0, 2326, 1386 ], "visible_bbox": [ 0, 0, 2326, 1386 ] }, { "name": "None", "role": "AXToolbar", "description": "None", "role_description": "toolbar", "value": "None", "children": [ { "name": "None", "role": "AXCheckBox", "description": "View", "role_description": "checkbox", "value": "0", "children": [], "bbox": [ 362, 0, 100, 104 ], "visible_bbox": [ 362, 0, 100, 104 ] }, { "name": "None", "role": "AXButton", "description": "Check", "role_description": "button", "value": "None", "children": [], "bbox": [ 970, 0, 94, 104 ], "visible_bbox": [ 970, 0, 94, 104 ] }, { "name": "None", "role": "AXCheckBox", "description": "Reader", "role_description": "checkbox", "value": "0", "children": [], "bbox": [ 1098, 0, 100, 104 ], "visible_bbox": [ 1098, 0, 100, 104 ] }, { "name": "None", "role": "AXButton", "description": "Slidehsow", "role_description": "button", "value": "None", "children": [], "bbox": [ 1198, 0, 100, 104 ], "visible_bbox": [ 1198, 0, 100, 104 ] }, { "name": "None", "role": "AXCheckBox", "description": "Star", "role_description": "checkbox", "value": "0", "children": [], "bbox": [ 1654, 0, 100, 104 ], "visible_bbox": [ 1654, 0, 100, 104 ] }, { "name": "None", "role": "AXButton", "description": "Share", "role_description": "button", "value": "None", "children": [], "bbox": [ 1754, 0, 100, 104 ], "visible_bbox": [ 1754, 0, 100, 104 ] }, { "name": "None", "role": "AXGroup", "description": "None", "role_description": "group", "value": "None", "children": [ { "name": "None", "role": "AXTextField", "description": "None", "role_description": "search text field", "value": "", "children": [ { "name": "", "role": "AXButton", "description": "search", "role_description": "button", "value": "None", "children": [], "bbox": [ 1864, 30, 50, 44 ], "visible_bbox": [ 1864, 30, 50, 44 ] } ], "bbox": [ 1860, 22, 444, 62 ], "visible_bbox": [ 1860, 22, 444, 62 ] } ], "bbox": [ 1854, 0, 456, 104 ], "visible_bbox": [ 1854, 0, 456, 104 ] } ], "bbox": [ 0, 0, 2326, 104 ], "visible_bbox": [ 0, 0, 2326, 104 ] }, { "name": "None", "role": "AXButton", "description": "None", "role_description": "close button", "value": "None", "children": [], "bbox": [ 38, 36, 28, 32 ], "visible_bbox": [ 38, 36, 28, 32 ] }, { "name": "None", "role": "AXButton", "description": "None", "role_description": "full screen button", "value": "None", "children": [ { "name": "None", "role": "AXGroup", "description": "None", "role_description": "group", "value": "None", "children": [ { "name": "None", "role": "AXGroup", "description": "None", "role_description": "group", "value": "None", "children": [], "bbox": [ 118, 36, 28, 32 ], "visible_bbox": [ 118, 36, 28, 32 ] } ], "bbox": [ 118, 36, 28, 32 ], "visible_bbox": [ 118, 36, 28, 32 ] } ], "bbox": [ 118, 36, 28, 32 ], "visible_bbox": [ 118, 36, 28, 32 ] }, { "name": "None", "role": "AXButton", "description": "None", "role_description": "minimize button", "value": "None", "children": [], "bbox": [ 78, 36, 28, 32 ], "visible_bbox": [ 78, 36, 28, 32 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "Machine Learning Blog | ML@CMU | Carnegie Mellon University", "children": [], "bbox": [ 496, 0, 426, 104 ], "visible_bbox": [ 496, 0, 426, 104 ] }, { "name": "None", "role": "AXStaticText", "description": "None", "role_description": "text", "value": "1 article", "children": [], "bbox": [ 496, 0, 426, 104 ], "visible_bbox": [ 496, 0, 426, 104 ] } ], "bbox": [ 0, 0, 2326, 1386 ], "visible_bbox": [ 0, 0, 2326, 1386 ] }