[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-122916-en":3,"doc-seo-122916-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},122916,137441390410,"Hazel","https://ap-avatar.wpscdn.com/avatar/2000252f4ab5702993?_k=1776741390130283984",8,"Research & Report","On The Fairness Impacts of Hardware Selection in Machine Learning - research paper","Machine learning hardware selection influences the balance between model performance and fairness, especially in ML-as-a-service settings where users have limited control over training and deployment hardware. The work shows hardware choices can amplify pre-existing disparities, driven by differences in gradient flows and loss surfaces across demographic groups. Using both theoretical reasoning and empirical validation, it pinpoints the mechanisms behind hardware-induced performance imbalance and offers a training-procedure-based mitigation strategy to reduce fairness degradation.","On The Fairness Impacts of Hardware Selection in Machine Learning  \nSree Harsha Nelaturu * 1 2 Nishaanth K Ravichandran * 1 Cuong Tran 3 4 Sara Hooker 5 Ferdinando Fioretto 4  \narXiv :2312 .03886v2 [ cs .LG] 30 Aug 2024  \nAbstract  \nIn the machine learning ecosystem hardware selection is often regarded as a mere utility, overshadowed by the spotlight on algorithms and data. This is especially relevant in contexts like machine learning as-a-service platforms, where users often lack control over the hardware used for model training and deployment. This paper investigates the influence of hardware on the delicate balance between model performance and fairness. We demonstrate that hardware choices can exacerbate existing disparities, and attribute these discrepancies to variations in gradient flows and loss surfaces across different demographic groups. Through both theoretical and empirical analysis, the paper not only identifies the underlying factors but also proposes an effective strategy for mitigating hardware-induced performance imbalances.  \n1. Introduction  \nThe leap in capabilities of modern machine learning (ML) models has been powered primarily by the availability of large-scale datasets, gains in available compute, and the development of algorithms that can effectively use these resources (Radford et al., 2019; Brown et al., 2020) . As MLbased systems become integral to decision-making processes that bear considerable social and economic consequences, questions about their ethical application inevitably surface. While an active area of research has been devoted to understanding algorithmic choices and their implications on fairness (Hooker et al., 2020; Quan et al., 2023; Caton & Haas, 2020) and robustness (Carlini & Wagner, 2017; Waqas et al., 2022) in neural networks, there has been limited work to date concerning the influence of hard-  \n*Equal contribution 1Cohere For AI Community 2 Saarland University 3Dyania Health 4University of Virginia 5Cohere For AI. Correspondence to: Sara Hooker \u003C[sarahooker@cohere.com](sarahooker@cohere.com) >, Ferdinando Fioretto \u003C[fioretto@virginia.edu](fioretto@virginia.edu) >.  \nProceedings of the 41 st International Conference on Machine Learning, Vienna, Austria. PMLR 235, 2024 . Copyright 2024 by the author(s) .  \nware tooling on these critical aspects of model performance (Hooker, 2020; Zhuang et al., 2022; Jean-Paul et al., 2019) .  \nThis inquiry is especially pertinent as the ML hardware landscape undergoes substantial diversification, from successive generations of GPUs to custom deep-learning accelerators like TPUs (Jouppi et al., 2017) .  \nThis is significant, as ML models are frequently trained on ML services where there is limited choice in hardware, often geared towards increasingly specialized AI hardware, encouraging the use of a narrow range of ML frameworks (Mince et al., 2023) . More importantly, recent studies have indicated that models trained on different hardware can exhibit varying levels of accuracy (Zhuang et al., 2022) . A potential explanation is that hardware-induced nuances, such as precision discrepancies and threading behaviours, may lead iterative optimizers to different local minima during training (Hooker, 2020) .  \nThis paper further shows that these hardware-induced variations can disproportionately affect different groups, leading to a “rich get richer, poor get poorer” dynamic. This effect is depicted in Figure 1, which shows the variable impact of hardware changes on both a facial recognition task accuracy (left) and on an image classification task (right) across demographic groups or classes. Note that the only variable factor in this study is the underlying GPU adopted for training, while all other sources of software-related randomness are controlled. Remarkably, while the accuracy rates for majority groups (illustrated with lighter colors) remain relatively stable across different hardware configurations, the rates for minority groups (darker c","cbCaia3riHnuzblN","https://ap.wps.com/l/cbCaia3riHnuzblN","pdf",996840,1,22,"English","en",105,"# Introduction\n## Motivation: hardware underemphasized in ML fairness\n## Hardware diversity and its potential to alter training outcomes\n# Related Work\n## Intersection of hardware selection and fairness in ML","[{\"question\":\"Why does hardware selection matter for fairness in machine learning?\",\"answer\":\"Hardware choices can change how optimization behaves across demographic groups, causing unequal performance outcomes. The paper attributes these effects to differences in gradient flows and loss surfaces.\"},{\"question\":\"What settings does the paper highlight as especially relevant?\",\"answer\":\"Machine learning as-a-service platforms are emphasized because users often lack control over the hardware used for training and deployment.\"},{\"question\":\"How does the paper explain the “rich get richer, poor get poorer” dynamic?\",\"answer\":\"When models are trained on different hardware, accuracy can remain stable for majority groups while minority groups experience larger variability. The paper links this to group-dependent gradient flow behavior and distinct local loss surface geometry.\"}]","On The Fairness Impacts of Hardware Selection in Machine Learning - research paper | PDF",1785813643,55,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"on-the-fairness-impacts-of-hardware-selection-in-machine-learning-research-paper","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/on-the-fairness-impacts-of-hardware-selection-in-machine-learning-research-paper/122916/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-04",true,{"@type":65,"interactionType":66,"userInteractionCount":4},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why does hardware selection matter for fairness in machine learning?","Question",{"text":75,"@type":76},"Hardware choices can change how optimization behaves across demographic groups, causing unequal performance outcomes. The paper attributes these effects to differences in gradient flows and loss surfaces.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What settings does the paper highlight as especially relevant?",{"text":80,"@type":76},"Machine learning as-a-service platforms are emphasized because users often lack control over the hardware used for training and deployment.",{"name":82,"@type":73,"acceptedAnswer":83},"How does the paper explain the “rich get richer, poor get poorer” dynamic?",{"text":84,"@type":76},"When models are trained on different hardware, accuracy can remain stable for majority groups while minority groups experience larger variability. The paper links this to group-dependent gradient flow behavior and distinct local loss surface geometry.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]