{"id":1090,"date":"2022-04-22T13:34:00","date_gmt":"2022-04-22T04:34:00","guid":{"rendered":"https:\/\/citadelai.wpengine.com\/?p=1090"},"modified":"2024-09-17T12:20:20","modified_gmt":"2024-09-17T03:20:20","slug":"toward-standardization-for-bias-identification-and-management-in-ai-2-3-nist-special-publication-1270","status":"publish","type":"post","link":"https:\/\/citadel-ai.com\/ja\/blog\/2022\/04\/22\/toward-standardization-for-bias-identification-and-management-in-ai-2-3-nist-special-publication-1270\/","title":{"rendered":"AI\u306e\u30d0\u30a4\u30a2\u30b9\u7279\u5b9a\u30fb\u7ba1\u7406\u306e\u305f\u3081\u306e\u6a19\u6e96\u5316\u306b\u5411\u3051\u3066\uff082\/3\uff09NIST Special Publication 1270"},"content":{"rendered":"\n
\n

\u306f\u3058\u3081\u306b<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\u5f0a\u793e\u3067\u306f\u7c73\u56fd\u56fd\u7acb\u6a19\u6e96\u6280\u8853\u7814\u7a76\u6240(NIST)<\/strong><\/a>\u306e\u8a31\u53ef\u3092\u5f97\u3066\u30012022\u5e743\u6708\u306b\u767a\u8868\u3055\u308c\u305f\u5831\u544a\u66f8 \u201c<\/strong>NIST Special Publication 1270: Towards a Standard for Identifying and Managing Bias in Artificial Intelligence<\/strong><\/a>\u201d<\/strong>\u300cAI\u306e\u30d0\u30a4\u30a2\u30b9\u306e\u7279\u5b9a\u3068\u7ba1\u7406\u306e\u305f\u3081\u306e\u6a19\u6e96\u5316\u306b\u5411\u3051\u3066\u300d<\/strong>\u306b\u3064\u3044\u3066\u3001\u305d\u306e\u30ad\u30fc\u30dd\u30a4\u30f3\u30c8\u3068\u601d\u308f\u308c\u308b\u3068\u3053\u308d\u3092\u629c\u7c8b\u3057\u90a6\u8a33\u306e\u4e0a\u30013\u56de\u306b\u5206\u3051\u3066\u30d6\u30ed\u30b0\u3067\u516c\u958b\u3059\u308b\u3053\u3068\u3068\u81f4\u3057\u307e\u3057\u305f\u3002\u4eca\u56de\u306f\u7b2c\u4e8c\u56de\u76ee\u3067\u3059\u3002<\/p>\n\n\n\n

\u90a6\u8a33\u4e26\u3073\u306b\u30d6\u30ed\u30b0\u3067\u306e\u8ee2\u8f09\u306e\u8a31\u53ef\u3092\u9802\u304d\u307e\u3057\u305fNIST\u306e\u3054\u539a\u610f\u306b\u6df1\u8b1d\u7533\u3057\u4e0a\u3052\u307e\u3059\u3068\u5171\u306b\u3001\u672c\u8a18\u4e8b\u304c\u3001\u4eca\u5f8c\u306e\u65e5\u672c\u306e\u793e\u4f1a\u3084\u4f01\u696d\u306b\u304a\u3051\u308bAI\u306e\u5b9f\u88c5\u306b\u304a\u3044\u3066\u3001\u5c11\u3057\u3067\u3082\u7686\u69d8\u306e\u304a\u5f79\u306b\u7acb\u3066\u308b\u3053\u3068\u3092\u9858\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u5c1a\u3001\u90a6\u8a33\u306b\u95a2\u308f\u308b\u629c\u7c8b\u7b87\u6240\u306e\u9078\u5b9a\u306f\u3001\u3042\u304f\u307e\u3067\u5f0a\u793e\u306e\u79c1\u898b\u30fb\u8cac\u4efb\u306b\u5247\u3063\u305f\u3082\u306e\u3067\u3042\u308a\u3001NIST\u306e\u610f\u56f3\u3092\u73fe\u3057\u305f\u3082\u306e\u3067\u306f\u3054\u3056\u3044\u307e\u305b\u3093\u3002<\/p>\n\n\n\n

Translated with permission courtesy of the National Institute of Standards and Technology (NIST), not an official US Government translation. All rights reserved, US Secretary of Commerce.<\/em><\/p>\n\n\n\n

\u3010NIST Special Publication 1270\u306e\u69cb\u6210\u3011<\/h2>\n\n\n\n

\u524d\u56de<\/strong><\/a>\u306b\u5f15\u304d\u7d9a\u304d\u3001\u7b2c\u4e8c\u56de\u306f\u4ee5\u4e0b\u69cb\u6210\u306e\u5185\u30012. AI\u30d0\u30a4\u30a2\u30b9\u306e\u985e\u578b\u5316\u306e\u5f8c\u534a\u304b\u30892\u7ae0\u306e\u7d42\u308f\u308a\u307e\u3067\u306b\u3064\u3044\u3066\u3001\u4e3b\u306a\u8ad6\u70b9\u3092\u53d6\u308a\u307e\u3068\u3081\u307e\u3059\u3002<\/p>\n\n\n\n

Executive Summary<\/strong><\/p>\n\n\n\n

1 Purpose and Scope <\/strong>\u76ee\u7684\u3068\u30b9\u30b3\u30fc\u30d7<\/p>\n\n\n\n

2 AI Bias: Context and Terminology <\/strong>AI\u30d0\u30a4\u30a2\u30b9\u306e\u985e\u578b\u5316<\/p>\n\n\n\n

3 AI Bias: Challenges and Guidance <\/strong>AI\u30d0\u30a4\u30a2\u30b9\uff1a\u8ab2\u984c\u3068\u6307\u91dd<\/p>\n\n\n\n

4. Conclusions <\/strong>\u7d50\u8ad6<\/p>\n\n\n\n

5. Glossary<\/strong><\/p>\n\n\n\n

\n

\u4ee5\u4e0b2.3\u9805\u3067\u8a18\u8f09\u3055\u308c\u3066\u3044\u308b\u3001\u793e\u4f1a\u6280\u8853\u7684\u306a\u30a2\u30d7\u30ed\u30fc\u30c1<\/strong>\u306f\u3001\u672c\u5831\u544a\u66f8\u3092\u901a\u3058\u3066NIST\u304c\u63d0\u8a00\u3059\u308b\u30ac\u30a4\u30c9\u30e9\u30a4\u30f3\u306e\u9aa8\u683c\u3092\u6210\u3059\u3082\u306e\u3067\u3059\u3002<\/p>\n<\/blockquote>\n\n\n\n

2 AI Bias: Context and Terminology <\/strong>AI\u30d0\u30a4\u30a2\u30b9\u306e\u985e\u578b\u5316<\/em><\/p>\n\n\n\n

2.3 A Socio-technical Systems Approach <\/strong>\u793e\u4f1a\u6280\u8853\u7684\u306a\u30a2\u30d7\u30ed\u30fc\u30c1<\/em><\/p>\n\n\n\n

Adopting a socio- technical perspective<\/strong> can enable a broader understanding of AI impacts and the key decisions that happen throughout, and beyond, the AI lifecycle\u2013such as whether technology is even a solution to a given task or problem [3, 109].<\/p>\n\n\n\n

\u793e\u4f1a\u6280\u8853\u7684\u306a\u30a2\u30d7\u30ed\u30fc\u30c1<\/strong>\u3092\u63a1\u7528\u3059\u308b\u3053\u3068\u3067\u3001AI\u306e\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u306a\u3089\u3073\u306b\u305d\u306e\u5148\u306b\u8d77\u3053\u308a\u5f97\u308b\u5f71\u97ff\u3092\u3088\u308a\u5e83\u304f\u7406\u89e3\u3057\u3001\u4e0e\u3048\u3089\u308c\u305f\u8ab2\u984c\u3084\u554f\u984c\u306b\u5bfe\u3057\u3066\u6280\u8853\u3060\u3051\u304c\u771f\u306e\u89e3\u6c7a\u7b56\u3068\u306a\u308b\u306e\u304b\u3068\u3044\u3063\u305f\u91cd\u8981\u306a\u5224\u65ad\u3092\u884c\u3046\u3053\u3068\u304c\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002 [3, 109]<\/p>\n\n\n\n

Reframing AI-related factors such as datasets, TEVV, participatory design, and human-in-the-loop practices through a socio- technical lens means understanding how they are both functions of society and, through the power of AI, can impact society.<\/p>\n\n\n\n

AI\u306b\u95a2\u9023\u3059\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3001TEVV(Test, Evaluation, Validation, and Verification)\u3001\u53c2\u52a0\u578b\u8a2d\u8a08\u3068\u3044\u3063\u305f\u8981\u7d20\u3068\u3001\u30d2\u30e5\u30fc\u30de\u30f3\u30a4\u30f3\u30b6\u30eb\u30fc\u30d7\u306e\u5b9f\u8df5\u3092\u3001\u793e\u4f1a\u6280\u8853\u7684\u306a\u30ec\u30f3\u30ba\u3092\u901a\u3057\u3066\u6349\u3048\u76f4\u3059\u3053\u3068\u3067\u3001\u305d\u308c\u3089\u304c\u5171\u306b\u793e\u4f1a\u306e\u91cd\u8981\u306a\u6a5f\u80fd\u3067\u3042\u308b\u3068\u540c\u6642\u306b\u3001\u9006\u306bAI\u3068\u3044\u3046\u529b\u3092\u901a\u3058\u3066\u793e\u4f1a\u306b\u5f71\u97ff\u3092\u53ca\u307c\u3057\u5f97\u308b\u3082\u306e\u3067\u3042\u308b\u304b\u3068\u3044\u3046\u3053\u3068\u3092\u7406\u89e3\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

2.4 An Updated AI Lifecycle <\/strong>\u65b0\u305f\u306aAI\u306e\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb<\/em><\/p>\n\n\n\n

This document has adapted a four-stage AI lifecycle<\/strong> from other stakeholder ver sions.11 The intent is to enable AI designers, developers, evaluators and deployers to relate lifecycle processes with AI bias categories and effectively facilitate its identification and management.<\/p>\n\n\n\n

\u672c\u5831\u544a\u66f8\u306f\u3001\u4ed6\u306e\u8457\u4f5c\u7248\u306b\u8a18\u8f09\u3055\u308c\u305f\u30014 \u6bb5\u968e\u306e AI \u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb<\/strong>\u3092\u63a1\u7528\u3057\u307e\u3059\u3002\u305d\u306e\u610f\u56f3\u306f\u3001AI\u306e\u8a2d\u8a08\u8005\u3001\u958b\u767a\u8005\u3001\u8a55\u4fa1\u8005\u3001\u30c7\u30d7\u30ed\u30a4\u62c5\u5f53\u8005\u304c\u3001\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u30d7\u30ed\u30bb\u30b9\u3068AI\u306e\u30d0\u30a4\u30a2\u30b9\u30ab\u30c6\u30b4\u30ea\u30fc\u3092\u95a2\u9023\u4ed8\u3051\u3001\u52b9\u7387\u7684\u306b\u305d\u306e\u7279\u5b9a\u3068\u7ba1\u7406\u3092\u4fc3\u9032\u3067\u304d\u308b\u3088\u3046\u306b\u3059\u308b\u305f\u3081\u3067\u3059\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

AI Lifecycles are iterative<\/strong>, and begin in the Pre-Design stage<\/strong>, where planning, problem specification, background research, and identification of data take place.<\/p>\n\n\n\n

AI\u306e\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u306f\u53cd\u5fa9\u7684<\/strong>\u3067\u3042\u308a\u3001\u305d\u3057\u3066\u305d\u308c\u306f\u8a08\u753b\u3092\u7acb\u3066\u3001\u8ab2\u984c\u3092\u8a2d\u5b9a\u3057\u3001\u80cc\u666f\u8abf\u67fb\u306e\u4e0a\u30c7\u30fc\u30bf\u3092\u7279\u5b9a\u3059\u308b\u4e8b\u524d\u8a2d\u8a08\u30b9\u30c6\u30fc\u30b8<\/strong>\u304b\u3089\u59cb\u307e\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

The Design and Development stage<\/strong> typically starts with analysis of the requirements and the available data. Based on this, a model is designed or selected.<\/p>\n\n\n\n

\u8a2d\u8a08\u3068\u958b\u767a\u30b9\u30c6\u30fc\u30b8<\/strong>\u3067\u306f\u3001\u901a\u5e38\u3001\u8981\u6c42\u5206\u6790\u3068\u5229\u7528\u53ef\u80fd\u306a\u30c7\u30fc\u30bf\u306e\u5206\u6790\u304b\u3089\u59cb\u307e\u308a\u307e\u3059\u3002\u3053\u308c\u306b\u57fa\u3065\u3044\u3066\u3001\u30e2\u30c7\u30eb\u304c\u8a2d\u8a08\u3055\u308c\u305f\u308a\u9078\u629e\u3055\u308c\u305f\u308a\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

The Deployment stage<\/strong> is when the AI system is released and used. Once humans begin to interact with the AI system the performance of the system must be monitored and reassessed to ensure proper function.<\/p>\n\n\n\n

\u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8\u30b9\u30c6\u30fc\u30b8<\/strong>\u306f\u3001AI\u30b7\u30b9\u30c6\u30e0\u304c\u30ea\u30ea\u30fc\u30b9\u3055\u308c\u3001\u4f7f\u7528\u3055\u308c\u308b\u6642\u3067\u3059\u3002\u4eba\u9593\u304cAI\u30b7\u30b9\u30c6\u30e0\u3092\u5229\u7528\u3057\u59cb\u3081\u305f\u3089\u3001\u9069\u5207\u306a\u6a5f\u80fd\u3092\u4fdd\u8a3c\u3059\u308b\u305f\u3081\u306b\u30b7\u30b9\u30c6\u30e0\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u76e3\u8996\u3057\u3001\u518d\u8a55\u4fa1\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

The Test and Evaluation stage<\/strong> is continuous throughout the entire AI Development Lifecycle. Organizations are encouraged to perform continuous testing and evaluation<\/strong> of all AI system components and features where bias can contribute to harmful impacts.<\/p>\n\n\n\n

\u30c6\u30b9\u30c8\u3068\u8a55\u4fa1\u30b9\u30c6\u30fc\u30b8<\/strong>\u306f\u3001AI\u958b\u767a\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u5168\u4f53\u3092\u901a\u3058\u3066\u7d99\u7d9a\u7684\u306b\u884c\u308f\u308c\u307e\u3059\u3002\u7d44\u7e54\u306b\u306f\u3001\u30d0\u30a4\u30a2\u30b9\u304c\u6709\u5bb3\u306a\u5f71\u97ff\u3092\u3082\u305f\u3089\u3059\u53ef\u80fd\u6027\u306e\u3042\u308b\u3059\u3079\u3066\u306eAI\u30b7\u30b9\u30c6\u30e0\u306e\u30b3\u30f3\u30dd\u30fc\u30cd\u30f3\u30c8\u3068\u6a5f\u80fd\u306b\u3064\u3044\u3066\u3001\u7d99\u7d9a\u7684\u306a\u30c6\u30b9\u30c8\u3068\u8a55\u4fa1<\/strong>\u3092\u884c\u3046\u3053\u3068\u304c\u63a8\u5968\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

For example, if during deployment the model is retrained with new data for a specific context<\/strong>, the model deployer should work with the model producer to assess actual performance for bias evaluation. Multi-stakeholder engagement is encouraged to ensure that the assessment is balanced and comprehensive.<\/p>\n\n\n\n

\u4f8b\u3048\u3070\u3001\u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8\u306e\u904e\u7a0b\u3067\u3001\u7279\u5b9a\u306e\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u306e\u65b0\u3057\u3044\u30c7\u30fc\u30bf\u3067\u30e2\u30c7\u30eb\u3092\u518d\u5b66\u7fd2\u3059\u308b\u5834\u5408<\/strong>\u3001\u30e2\u30c7\u30eb\u306e\u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8\u62c5\u5f53\u8005\u306f\u30e2\u30c7\u30eb\u958b\u767a\u8005\u3068\u5354\u529b\u3057\u3066\u3001\u30d0\u30a4\u30a2\u30b9\u8a55\u4fa1\u306e\u305f\u3081\u306b\u5b9f\u969b\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u691c\u8a3c\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002\u691c\u8a3c\u304c\u30d0\u30e9\u30f3\u30b9\u306e\u3068\u308c\u305f\u5305\u62ec\u7684\u306a\u3082\u306e\u3067\u3042\u308b\u3053\u3068\u3092\u4fdd\u8a3c\u3059\u308b\u305f\u3081\u306b\u306f\u3001\u3055\u307e\u3056\u307e\u306a\u30b9\u30c6\u30fc\u30af\u30db\u30eb\u30c0\u30fc\u304c\u691c\u8a3c\u306b\u95a2\u4e0e\u3059\u308b\u3053\u3068\u304c\u63a8\u5968\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

If deviations from desired goals are observed, the findings should feed into the model Pre-Design stage<\/strong> to ensure appropriate adjustments are made in data curation and problem formulation.<\/p>\n\n\n\n

\u671b\u307e\u3057\u3044\u76ee\u6a19\u5024\u304b\u3089\u5916\u308c\u3066\u3044\u308b\u3053\u3068\u304c\u89b3\u6e2c\u3055\u308c\u305f\u5834\u5408\u3001\u305d\u306e\u77e5\u898b\u3092\u30e2\u30c7\u30eb\u306e\u4e8b\u524d\u8a2d\u8a08\u30b9\u30c6\u30fc\u30b8<\/strong>\u306b\u53cd\u6620\u3055\u305b\u3001\u30c7\u30fc\u30bf\u53ce\u96c6\u3001\u53ca\u3073\u8ab2\u984c\u8a2d\u5b9a\u306b\u304a\u3044\u3066\u9069\u5207\u306a\u8abf\u6574\u304c\u306a\u3055\u308c\u308b\u3088\u3046\u306b\u3059\u3079\u304d\u3067\u3059\u3002<\/p>\n\n\n\n

Any proposed changes to the design of the model should then be evaluated together with the new data and requirements to ensure compatibility and identification of any potential new sources of bias<\/strong>.<\/p>\n\n\n\n

\u30e2\u30c7\u30eb\u8a2d\u8a08\u306b\u63d0\u6848\u3055\u308c\u305f\u3059\u3079\u3066\u306e\u5909\u66f4\u306f\u3001\u65b0\u3057\u3044\u30c7\u30fc\u30bf\u53ca\u3073\u8981\u6c42\u6761\u4ef6\u3068\u5408\u308f\u305b\u3066\u8a55\u4fa1\u306e\u4e0a\u3001\u4e92\u63db\u6027\u3092\u691c\u8a3c\u3057\u3001\u65b0\u305f\u306a\u30d0\u30a4\u30a2\u30b9\u306e\u539f\u56e0\u3068\u306a\u308a\u5f97\u308b\u3082\u306e\u3092\u6d17\u3044\u51fa\u3059\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059<\/strong>\u3002<\/p>\n\n\n\n

Then another round of design and implementation<\/strong> commences to formulate corresponding requirements for the new model capabilities and features and for additional datasets. During this stage, the model developer should perform continuous testing and evaluation<\/strong> to ensure that bias mitigation maintains effectiveness in the new setting, as the model is optimized and tested for performance.<\/p>\n\n\n\n

\u6b21\u306b\u3001\u65b0\u3057\u3044\u30e2\u30c7\u30eb\u306e\u6a5f\u80fd\u3068\u7279\u5fb4\u91cf\u3001\u4e26\u3073\u306b\u8ffd\u52a0\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u5bfe\u5fdc\u3059\u308b\u8981\u4ef6\u3092\u7b56\u5b9a\u3059\u308b\u3001\u65b0\u305f\u306a\u8a2d\u8a08\u30fb\u5b9f\u88c5\u30b9\u30c6\u30fc\u30b8<\/strong>\u306b\u5165\u308a\u307e\u3059\u3002\u3053\u306e\u30b9\u30c6\u30fc\u30b8\u3067\u306f\u3001\u30e2\u30c7\u30eb\u306e\u6700\u9069\u5316\u3068\u6027\u80fd\u30c6\u30b9\u30c8\u3092\u884c\u306a\u3044\u306a\u304c\u3089\u3001\u30d0\u30a4\u30a2\u30b9\u8efd\u6e1b\u304c\u65b0\u3057\u3044\u74b0\u5883\u3067\u3082\u6709\u52b9\u6027\u3092\u7dad\u6301\u3067\u304d\u308b\u3088\u3046\u3001\u30e2\u30c7\u30eb\u958b\u767a\u8005\u306f\u7d99\u7d9a\u7684\u306b\u30c6\u30b9\u30c8\u3068\u8a55\u4fa1\u3092\u5b9f\u65bd<\/strong>\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

Once released, the deploying organization should use documented model specifications to test and evaluate bias characteristics during deployment in the specific context. Ideally, this evaluation should be performed together with other stakeholders to ensure all previously identified problems are resolved to everyone\u2019s satisfaction.<\/p>\n\n\n\n

\u30ea\u30ea\u30fc\u30b9\u5f8c\u3001\u30c7\u30d7\u30ed\u30a4\u62c5\u5f53\u90e8\u9580\u306f\u6587\u66f8\u5316\u3055\u308c\u305f\u30e2\u30c7\u30eb\u4ed5\u69d8\u3092\u7528\u3044\u3066\u3001\u5404\u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8\u74b0\u5883\u4e0b\u306b\u304a\u3051\u308b\u30d0\u30a4\u30a2\u30b9\u7279\u6027\u3092\u30c6\u30b9\u30c8\u3057\u8a55\u4fa1\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002\u7406\u60f3\u7684\u306b\u306f\u3001\u3053\u306e\u8a55\u4fa1\u306f\u4ed6\u306e\u30b9\u30c6\u30fc\u30af\u30db\u30eb\u30c0\u30fc\u3068\u3068\u3082\u306b\u5b9f\u65bd\u3057\u3001\u4e8b\u524d\u306b\u7279\u5b9a\u3055\u308c\u305f\u554f\u984c\u304c\u3059\u3079\u3066\u89e3\u6c7a\u3055\u308c\u3001\u5168\u54e1\u304c\u6e80\u8db3\u3059\u308b\u3088\u3046\u306b\u3059\u3079\u304d\u3067\u3059\u3002<\/p>\n\n\n\n

3. AI Bias: Challenges and Guidance <\/strong>AI \u30d0\u30a4\u30a2\u30b9\uff1a\u8ab2\u984c\u3068\u6307\u91dd<\/em><\/p>\n\n\n\n

Through a review of the literature, and various multi-stakeholder processes, including public comments, workshops, and listening sessions, NIST has identified three broad areas that present challenges for addressing AI bias.<\/p>\n\n\n\n

\u6587\u732e\u306e\u30ec\u30d3\u30e5\u30fc\u3084\u3001\u30d1\u30d6\u30ea\u30c3\u30af\u30b3\u30e1\u30f3\u30c8\u3001\u30ef\u30fc\u30af\u30b7\u30e7\u30c3\u30d7\u3001\u30ea\u30b9\u30cb\u30f3\u30b0\u30bb\u30c3\u30b7\u30e7\u30f3\u306a\u3069\u306e\u69d8\u3005\u306a\u30b9\u30c6\u30fc\u30af\u30db\u30eb\u30c0\u30fc\u3068\u306e\u30d7\u30ed\u30bb\u30b9\u3092\u901a\u3058\u3066\u3001NIST\u306f\u3001AI\u306e\u30d0\u30a4\u30a2\u30b9\u306b\u5bfe\u51e6\u3059\u308b\u305f\u3081\u306e\u8ab2\u984c\u3068\u3057\u3066\u30013\u3064\u306e\u5e83\u7bc4\u306a\u9818\u57df<\/strong>\u3092\u660e\u3089\u304b\u306b\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

The first challenge relates to dataset factors<\/strong> such as availability, representativeness, and baked-in societal biases.<\/p>\n\n\n\n

The second relates to issues of measurement and metrics<\/strong> to support testing and evaluation, validation, and verification (TEVV).<\/p>\n\n\n\n

The third area broadly comprises issues related to human factors<\/strong>, including societal and historic biases within individuals and organizations, as well as challenges related to implementing human-in-the-loop.<\/p>\n\n\n\n

\u6700\u521d\u306e\u8ab2\u984c\u306f\u3001\u5229\u7528\u53ef\u80fd\u6027\u3001\u4ee3\u8868\u6027\u3001\u793e\u4f1a\u7684\u306a\u30d0\u30a4\u30a2\u30b9\u306a\u3069\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8<\/strong>\u306b\u95a2\u3059\u308b\u3082\u306e\u3067\u3059\u3002<\/p>\n\n\n\n

2\u3064\u76ee\u306f\u3001\u30c6\u30b9\u30c8\u3068\u8a55\u4fa1\u3001\u691c\u8a3c\u3001\u59a5\u5f53\u6027\u78ba\u8a8d\uff08TEVV\uff09\u3092\u30b5\u30dd\u30fc\u30c8\u3059\u308b\u305f\u3081\u306e\u6e2c\u5b9a\u3068\u30e1\u30c8\u30ea\u30c3\u30af\u306e\u554f\u984c<\/strong>\u306b\u95a2\u3059\u308b\u3082\u306e\u3067\u3059\u3002<\/p>\n\n\n\n

3\u3064\u76ee\u306e\u5206\u91ce\u306f\u3001\u500b\u4eba\u3084\u7d44\u7e54\u306b\u304a\u3051\u308b\u793e\u4f1a\u7684\u30fb\u6b74\u53f2\u7684\u30d0\u30a4\u30a2\u30b9\u3084\u3001\u30d2\u30e5\u30fc\u30de\u30f3\u30a4\u30f3\u30b6\u30eb\u30fc\u30d7\u306e\u5b9f\u88c5\u306b\u95a2\u3059\u308b\u8ab2\u984c\u306a\u3069\u3001\u4eba\u7684\u8981\u56e0<\/strong>\u306b\u95a2\u9023\u3059\u308b\u554f\u984c\u3092\u5e83\u304f\u542b\u3093\u3067\u3044\u307e\u3059<\/p>\n\n\n\n

NIST plans to work with the trustworthy and responsible AI communities to explore the proposed mitigants and governance processes, and build associated formal technical guidance over the coming years in concert with these communities.<\/p>\n\n\n\n

NIST\u306f\u3001\u4fe1\u983c\u3067\u304d\u308b\u8cac\u4efb\u3042\u308bAI\u306e\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3068\u5354\u529b\u3057\u3066\u3001\u63d0\u6848\u3055\u308c\u3066\u3044\u308b\u7de9\u548c\u7b56\u3084\u30ac\u30d0\u30ca\u30f3\u30b9\u30d7\u30ed\u30bb\u30b9\u3092\u691c\u8a0e\u3057\u3001\u3053\u308c\u3089\u306e\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3068\u9023\u643a\u3057\u3066\u3001\u4eca\u5f8c\u6570\u5e74\u9593\u3067\u95a2\u9023\u3059\u308b\u6b63\u5f0f\u306a\u6280\u8853\u30ac\u30a4\u30c0\u30f3\u30b9\u3092\u69cb\u7bc9\u3059\u308b\u4e88\u5b9a\u3067\u3059\u3002<\/p>\n\n\n\n

3.1 Who is Counted? Datasets in AI Bias<\/strong><\/p>\n\n\n\n

\u8ab0\u3092\u8003\u616e\u3057\u4f55\u3092\u8003\u616e\u3057\u306a\u3044\u306e\u304b\uff1fAI\u30d0\u30a4\u30a2\u30b9\u306b\u304a\u3051\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u610f\u5473<\/em><\/p>\n\n\n\n

3.1.1 Dataset Challenges <\/strong>\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u8ab2\u984c<\/em><\/p>\n\n\n\n

AI design and development practices rely on large scale datasets to drive ML processes. This ever-present need can lead researchers, developers, and practitioners to first \u201cgo where the data is,\u201d and adapt their questions accordingly [124]. This creates a culture focused more on which datasets are available or accessible<\/strong>, rather than what dataset might be most suitable [108]<\/p>\n\n\n\n

AI\u306e\u8a2d\u8a08\u3068\u958b\u767a\u306e\u73fe\u5834\u3067\u306f\u3001\u6a5f\u68b0\u5b66\u7fd2\u30d7\u30ed\u30bb\u30b9\u3092\u5b9f\u884c\u3059\u308b\u306b\u3042\u305f\u308a\u3001\u5927\u898f\u6a21\u306a\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u304c\u5fc5\u8981\u3067\u3059\u3002\u3053\u306e\u7d76\u3048\u305a\u5b58\u5728\u3059\u308b\u30cb\u30fc\u30ba\u306f\u3001 AI\u306e\u7814\u7a76\u8005\u3001\u958b\u767a\u8005\u3001\u5b9f\u52d9\u5bb6\u3092\u3001\u30c7\u30fc\u30bf\u304c\u3042\u308b\u3068\u3053\u308d\u306b\u99c6\u308a\u7acb\u3066\u3001\u8ab2\u984c\u3092\u305d\u308c\u306b\u9069\u5408\u3055\u305b\u3066\u3057\u307e\u3046\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059[124]\u3002\u3053\u306e\u3053\u3068\u306f\u3001\u3069\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u304c\u6700\u3082\u9069\u3057\u3066\u3044\u308b\u304b\u3068\u3044\u3046\u3053\u3068\u3088\u308a\u3082\u3001\u3069\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u304c\u5229\u7528\u53ef\u80fd\u304b\u3001\u30a2\u30af\u30bb\u30b9\u53ef\u80fd\u304b\u3068\u3044\u3046\u3053\u3068\u306b\u3001\u3088\u308a\u7126\u70b9\u3092\u5f53\u3066\u308b\u6587\u5316\u3092\u751f\u307f\u51fa\u3059<\/strong>\u3053\u3068\u306b\u306a\u308a\u307e\u3059[108]\u3002<\/p>\n\n\n\n

As a result, the data used in these processes may not be fully representative of populations or the phenomena that are being modeled. The data that is collected can differ significantly from what occurs in the real world<\/strong> [77, 78, 119].<\/p>\n\n\n\n

\u305d\u306e\u7d50\u679c\u3001\u3053\u308c\u3089\u306e\u30d7\u30ed\u30bb\u30b9\u3067\u4f7f\u7528\u3055\u308c\u308b\u30c7\u30fc\u30bf\u306f\u3001\u6bcd\u96c6\u56e3\u3084\u30e2\u30c7\u30eb\u5316\u3055\u308c\u308b\u73fe\u8c61\u3092\u5fc5\u305a\u3057\u3082\u4ee3\u8868\u3057\u3066\u3044\u308b\u3068\u306f\u8a00\u3044\u96e3\u3044\u3082\u306e\u3068\u306a\u308b\u304b\u3082\u77e5\u308c\u307e\u305b\u3093\u3002\u53ce\u96c6\u3055\u308c\u305f\u30c7\u30fc\u30bf\u304c\u3001\u5b9f\u4e16\u754c\u3067\u8d77\u3053\u3063\u3066\u3044\u308b\u3053\u3068\u3068\u5927\u304d\u304f\u7570\u306a\u3063\u3066\u3057\u307e\u3046\u53ef\u80fd\u6027\u3082\u3042\u308a\u5f97\u307e\u3059<\/strong> [77, 78, 119]\u3002<\/p>\n\n\n\n

Systemic biases<\/strong> may also be manifested in the form of availability bias<\/strong> when datasets that are readily available but not fully representative of the target population (including proxy data) are used and reused as training data.<\/p>\n\n\n\n

\u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9<\/strong>\uff08\u8a33\u6ce8\uff1a\u7d44\u7e54\u5168\u4f53\u306b\u53ca\u3076\u3088\u3046\u306a\u30d0\u30a4\u30a2\u30b9\uff09\u304c\u9855\u5728\u5316\u3059\u308b\u4e00\u4f8b\u3068\u3057\u3066\u306f\u3001\u5165\u624b\u3057\u6613\u3044\u4e00\u65b9\u3067\u3001\u5bfe\u8c61\u6bcd\u96c6\u56e3\u3092\u5341\u5206\u306b\u4ee3\u8868\u3057\u3066\u3044\u306a\u3044\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\uff08\u30d7\u30ed\u30ad\u30b7\u30c7\u30fc\u30bf\u3092\u542b\u3080\uff09\u3092\u5b66\u7fd2\u30c7\u30fc\u30bf\u3068\u3057\u3066\u4f7f\u7528\u30fb\u518d\u5229\u7528\u3057\u3066\u3057\u307e\u3063\u305f\u5834\u5408\u306b\u767a\u751f\u3059\u308b\u3001\u5229\u7528\u53ef\u80fd\u6027\u30d0\u30a4\u30a2\u30b9<\/strong>\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

Other issues arise due to the common ML practice of reusing datasets<\/strong>. Under such practices, datasets may become disconnected from the social contexts and time periods of their creation<\/strong>.<\/p>\n\n\n\n

\u6a5f\u68b0\u5b66\u7fd2\u306e\u5b9f\u8df5\u74b0\u5883\u3067\u306f\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3092\u518d\u5229\u7528<\/strong>\u3059\u308b\u3053\u3068\u304c\u591a\u3044\u306e\u3067\u3059\u304c\u3001\u305d\u3046\u3057\u305f\u624b\u6cd5\u306b\u8d77\u56e0\u3059\u308b\u554f\u984c\u3082\u3042\u308a\u307e\u3059\u3002\u3053\u306e\u3088\u3046\u306a\u30c7\u30fc\u30bf\u518d\u5229\u7528\u306e\u624b\u6cd5\u3092\u7528\u3044\u308b\u5834\u5408\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u304c\u3001\u305d\u308c\u304c\u4f5c\u6210\u3055\u308c\u305f\u793e\u4f1a\u7684\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3084\u6642\u4ee3\u304b\u3089\u5207\u308a\u96e2\u3055\u308c\u3066\u3057\u307e\u3044\u307e\u3059<\/strong>\u3002<\/p>\n\n\n\n

Even when datasets are representative, they may still exhibit entrenched historical and systemic biases, improperly utilize protected attributes, or utilize culturally or contextually unsuitable attributes<\/strong>. Latent variables such as gender can be inferred through browsing history, and race can be inferred through zip code.<\/p>\n\n\n\n

\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u304c\u6bcd\u96c6\u56e3\u3092\u4ee3\u8868\u3057\u3066\u3044\u305f\u3068\u3057\u3066\u3082\u3001\u6b74\u53f2\u7684\u30d0\u30a4\u30a2\u30b9\u3084\u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9\u304c\u6b8b\u3063\u3066\u3044\u305f\u308a\u3001\u672c\u6765\u5229\u7528\u3059\u3079\u304d\u3067\u306f\u306a\u3044\u30d7\u30ed\u30c6\u30af\u30c8\u3055\u308c\u305f\u5c5e\u6027\u3084\u3001\u6587\u5316\u7684\u30fb\u6587\u8108\u7684\u306b\u4e0d\u9069\u5207\u306a\u5c5e\u6027\u304c\u5229\u7528\u3055\u308c\u305f\u308a\u3059\u308b<\/strong>\u3053\u3068\u304c\u3042\u308a\u307e\u3059\u3002\u4f8b\u3048\u3070\u3001\u6027\u5225\u3068\u3044\u3063\u305f\u6f5c\u5728\u5909\u6570\u306f\u95b2\u89a7\u5c65\u6b74\u304b\u3089\u3001\u4eba\u7a2e\u306f\u90f5\u4fbf\u756a\u53f7\u304b\u3089\u63a8\u6e2c\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u3067\u3059\u3002<\/p>\n\n\n\n

3.1.2 Dataset Guidance <\/strong>\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u30ac\u30a4\u30c0\u30f3\u30b9<\/em><\/p>\n\n\n\n

Not only is the predictive behavior of the ML system determined by the data, but the data also largely defines the machine learning task itself <\/strong>[62].<\/p>\n\n\n\n

\u30c7\u30fc\u30bf\u306f\u3001\u5358\u306b\u6a5f\u68b0\u5b66\u7fd2\u30b7\u30b9\u30c6\u30e0\u306e\u4e88\u6e2c\u7d50\u679c\u3092\u5de6\u53f3\u3059\u308b\u3068\u3044\u3046\u3060\u3051\u3067\u306a\u304f\u3001\u30c7\u30fc\u30bf\u306b\u306f\u3001\u6a5f\u68b0\u5b66\u7fd2\u306e\u30bf\u30b9\u30af\u305d\u306e\u3082\u306e\u3092\u5927\u304d\u304f\u6c7a\u5b9a\u4ed8\u3051\u3066\u3057\u307e\u3046\u3068\u3044\u3046\u5074\u9762\u304c\u3042\u308a\u307e\u3059\u3002<\/strong>[62]<\/p>\n\n\n\n

The question of dataset fit or suitability requires attention to three factors: statistical methods<\/strong> for mitigating representation issues; processes to account for the socio-technical context <\/strong>in which the application is being deployed; and awareness of the interaction of human factors<\/strong> with the AI technical system at all stages of the AI lifecycle.<\/p>\n\n\n\n

\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u9069\u5408\u6027\u307e\u305f\u306f\u9069\u6027\u306e\u554f\u984c\u306b\u53d6\u308a\u7d44\u3080\u306b\u306f\u30013\u3064\u306e\u8981\u56e0\u306b\u6ce8\u610f\u3092\u6255\u3046\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\uff1a<\/p>\n\n\n\n

\u4ee3\u8868\u6027\u306e\u554f\u984c\u3092\u8efd\u6e1b\u3059\u308b\u305f\u3081\u306e\u7d71\u8a08\u7684\u624b\u6cd5<\/strong>\u3001\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u304c\u30c7\u30d7\u30ed\u30a4\u3055\u308c\u308b\u793e\u4f1a\u6280\u8853\u7684\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8<\/strong>\u3092\u8003\u616e\u3059\u308b\u30d7\u30ed\u30bb\u30b9\u3001\u304a\u3088\u3073\u5168\u3066\u306eAI\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u30b9\u30c6\u30fc\u30b8\u306b\u304a\u3044\u3066\u3001\u4eba\u7684\u8981\u56e0\u3068AI\u6280\u8853\u30b7\u30b9\u30c6\u30e0\u3068\u306e\u76f8\u4e92\u4f5c\u7528<\/strong>\u3092\u8a8d\u8b58\u3059\u308b\u3053\u3068\u3067\u3059\u3002<\/p>\n\n\n\n

\n

Statistical Factors <\/strong>\u7d71\u8a08\u7684\u8981\u56e0<\/p>\n<\/blockquote>\n\n\n\n

AI bias problems are exacerbated by the variety of statistical biases that are prevalent in the large scale datasets used in ML modeling. When these models are deployed for decision-based applications, often in high-risk settings and off-label uses<\/strong>, harms can be perpetuated and amplified.<\/p>\n\n\n\n

AI\u30d0\u30a4\u30a2\u30b9\u306e\u554f\u984c\u306f\u3001\u6a5f\u68b0\u5b66\u7fd2\u30e2\u30c7\u30ea\u30f3\u30b0\u306b\u4f7f\u7528\u3055\u308c\u308b\u5927\u898f\u6a21\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u904d\u5728\u3059\u308b\u3001\u69d8\u3005\u306a\u7d71\u8a08\u7684\u30d0\u30a4\u30a2\u30b9\u306b\u3088\u3063\u3066\u3055\u3089\u306b\u6df1\u523b\u5316\u3057\u307e\u3059\u3002\u3053\u308c\u3089\u306e\u30e2\u30c7\u30eb\u304c\u3001\u610f\u601d\u6c7a\u5b9a\u306b\u95a2\u308f\u308b\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306b\u30c7\u30d7\u30ed\u30a4\u3055\u308c\u305f\u5834\u5408\u3001\u7279\u306b\u9ad8\u30ea\u30b9\u30af\u3067\u9069\u5fdc\u7bc4\u56f2\u5916\u3067\u4f7f\u7528\u3055\u308c\u305f\u5834\u5408<\/strong>\u3001\u5f0a\u5bb3\u304c\u6c38\u7d9a\u3057\u5897\u5e45\u3055\u308c\u308b\u30ea\u30b9\u30af\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

Consequently, a model trained on biased and erroneous data may lead to biased and inaccurate predictions<\/strong>. Moreover, training a model on one dataset and using it to operate on another requires special care to account for potential differences in the distributions of the datasets that may further exacerbate the unfairness and errors of the model.<\/strong><\/p>\n\n\n\n

\u305d\u306e\u7d50\u679c\u3001\u504f\u3063\u305f\u8aa4\u3063\u305f\u30c7\u30fc\u30bf\u3067\u8a13\u7df4\u3055\u308c\u305f\u30e2\u30c7\u30eb\u306f\u3001\u504f\u3063\u305f\u4e0d\u6b63\u78ba\u306a\u4e88\u6e2c\u306b\u3064\u306a\u304c\u308a\u307e\u3059\u3002<\/strong>\u3055\u3089\u306b\u3001\u3042\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3067\u5b66\u7fd2\u3055\u305b\u305f\u30e2\u30c7\u30eb\u3092\u5225\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3067\u4f7f\u7528\u3059\u308b\u5834\u5408\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u9593\u306e\u6f5c\u5728\u7684\u306a\u5206\u5e03\u306e\u5dee\u7570\u3092\u8003\u616e\u3057\u3066\u3001\u7279\u5225\u306a\u6ce8\u610f\u3092\u6255\u3046\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002<\/strong>\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u304c\u3001\u30e2\u30c7\u30eb\u306e\u4e0d\u516c\u5e73\u611f\u3084\u8aa4\u5dee\u3092\u3055\u3089\u306b\u60aa\u5316\u3055\u305b\u308b\u53ef\u80fd\u6027\u304c\u3042\u308b\u304b\u3089\u3067\u3059\u3002<\/p>\n\n\n\n

\n

Accounting for Socio-technical Factors <\/strong>\u793e\u4f1a\u6280\u8853\u7684\u8981\u56e0\u306e\u8003\u616e<\/p>\n<\/blockquote>\n\n\n\n

While statistical methods are indeed necessary, they are not sufficient for addressing the AI bias challenges associated with datasets. Modeling processes have the intent of making contextual concepts measurable. Once the context has been removed, however, it is difficult to get it back, leading AI models to learn from inexact representations.<\/p>\n\n\n\n

\u7d71\u8a08\u7684\u624b\u6cd5\u306f\u78ba\u304b\u306b\u5fc5\u8981\u3067\u3059\u304c\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u95a2\u9023\u3059\u308bAI\u30d0\u30a4\u30a2\u30b9\u306e\u8ab2\u984c\u3092\u89e3\u6c7a\u3059\u308b\u306b\u306f\u5341\u5206\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u30e2\u30c7\u30ea\u30f3\u30b0\u30d7\u30ed\u30bb\u30b9\u306b\u304a\u3044\u3066\u306f\u3001\u30b3\u30f3\u30c6\u30af\u30b9\u30c8\u3092\u8003\u616e\u3057\u305f\u6982\u5ff5\u3092\u53ef\u8996\u5316\u3059\u308b\u3053\u3068\u3092\u76ee\u6307\u3057\u3066\u3044\u307e\u3059\u3002\u3057\u304b\u3057\u3001\u4e00\u65e6\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u304b\u3089\u96e2\u308c\u3066\u3057\u307e\u3046\u3068\u3001\u305d\u308c\u3092\u5143\u306b\u623b\u3059\u3053\u3068\u306f\u96e3\u3057\u304f\u3001\u6b63\u78ba\u306b\u9078\u3070\u308c\u305f\u3068\u306f\u8a00\u3044\u96e3\u3044\u30c7\u30fc\u30bf\u3092AI\u30e2\u30c7\u30eb\u304c\u5b66\u7fd2\u3057\u3066\u3057\u307e\u3046\u3053\u3068\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

The practice of deploying AI in off-label uses<\/strong>, that is AI systems being applied to a task or within a social or organizational context for which it was not designed, <\/strong>must be approached with caution, especially in high-risk settings.<\/p>\n\n\n\n

AI\u3092\u9069\u5fdc\u7bc4\u56f2\u5916\u3067\u30c7\u30d7\u30ed\u30a4\u3059\u308b\u3053\u3068\u3001\u3059\u306a\u308f\u3061\u3001\u610f\u56f3\u3055\u308c\u3066\u3044\u306a\u304b\u3063\u305f\u30bf\u30b9\u30af\u3084\u793e\u4f1a\u7684\u30fb\u7d44\u7e54\u7684\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u306bAI\u30b7\u30b9\u30c6\u30e0\u3092\u9069\u7528\u3059\u308b\u3053\u3068<\/strong>\u306f\u3001\u7279\u306b\u30ea\u30b9\u30af\u306e\u9ad8\u3044\u5834\u6240\u3067\u5229\u7528\u3055\u308c\u308b\u5834\u5408\u3001\u614e\u91cd\u306b\u53d6\u308a\u7d44\u307e\u306a\u3051\u308c\u3070\u306a\u308a\u307e\u305b\u3093\u3002<\/p>\n\n\n\n

\n

Interaction of human factors and datasets<\/strong><\/p>\n\n\n\n

\u4eba\u7684\u8981\u56e0\u3068\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u76f8\u4e92\u4f5c\u7528<\/p>\n<\/blockquote>\n\n\n\n

Systemic institutional biases<\/strong> are captured in the datasets used to build the models underlying AI applications. These biases are compounded by the decisions and assumptions made by AI design and development teams about which datasets to use [145]. These decisions affect who and what gets counted, and who and what does not get counted.<\/p>\n\n\n\n

AI\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306e\u57fa\u790e\u3068\u306a\u308b\u30e2\u30c7\u30eb\u3092\u69cb\u7bc9\u3059\u308b\u305f\u3081\u306b\u4f7f\u7528\u3055\u308c\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u306f\u3001\u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9<\/strong>\u304c\u5b58\u5728\u3057\u307e\u3059\u3002\u3053\u308c\u3089\u306e\u30d0\u30a4\u30a2\u30b9\u306f\u3001AI\u306e\u8a2d\u8a08\u30fb\u958b\u767a\u30c1\u30fc\u30e0\u304c\u3001\u3069\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3092\u4f7f\u7528\u3059\u308b\u304b\u5224\u65ad\u3057\u4eee\u5b9a\u3059\u308b\u3053\u3068\u306b\u3088\u3063\u3066\u9f4e\u3055\u308c\u307e\u3059[145]\u3002\u3053\u3046\u3057\u305f\u5224\u65ad\u304c\u3001\u8ab0\u3092\u8003\u616e\u3057\u3001\u4f55\u3092\u8003\u616e\u3057\u306a\u3044\u306e\u304b\u3068\u3044\u3063\u305f\u3053\u3068\u306b\u5f71\u97ff\u3092\u53ca\u307c\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

Data typically needs to be cleaned in some way, removing outliers and spurious data. Missing data may be imputed (replacing the missing values with nearest neighbors or extrapolated values) or removed entirely. Missing data may be more frequent<\/strong> in marginalized populations. Furthermore, because of compounding collection biases, missing and spurious data is often not random.<\/p>\n\n\n\n

\u30c7\u30fc\u30bf\u306f\u901a\u5e38\u3001\u4f55\u3089\u304b\u306e\u65b9\u6cd5\u3067\u30af\u30ec\u30f3\u30b8\u30f3\u30b0\u3055\u308c\u3001\u5916\u308c\u5024\u3084\u7591\u308f\u3057\u3044\u30c7\u30fc\u30bf\u3092\u9664\u53bb\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002\u6b20\u640d\u30c7\u30fc\u30bf\u306f\u3001\u30a4\u30f3\u30d4\u30e5\u30c6\u30fc\u30b7\u30e7\u30f3\uff08\u6b20\u640d\u5024\u3092\u6700\u8fd1\u508d\u5024\u307e\u305f\u306f\u5916\u633f\u5024\u3067\u7f6e\u304d\u63db\u3048\u308b\u3053\u3068\uff09\u307e\u305f\u306f\u5b8c\u5168\u306b\u9664\u53bb\u3055\u308c\u308b\u3053\u3068\u304c\u3042\u308a\u307e\u3059\u3002\u4e00\u65b9\u3067\u3001\u793e\u4f1a\u304b\u3089\u758e\u5916\u3055\u308c\u305f\u96c6\u56e3\u3067\u306f\u3001\u6b20\u640d\u3055\u308c\u305f\u3088\u3046\u306a\u30c7\u30fc\u30bf\u304c\u3088\u308a\u983b\u7e41\u306b\u767a\u751f\u3059\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059<\/strong>\u3002\u3055\u3089\u306b\u3001\u30c7\u30fc\u30bf\u53ce\u96c6\u306e\u969b\u306e\u30d0\u30a4\u30a2\u30b9\u304c\u8907\u5408\u7684\u306b\u4f5c\u7528\u3059\u308b\u305f\u3081\u3001\u6b20\u640d\u30c7\u30fc\u30bf\u3084\u7591\u308f\u3057\u3044\u30c7\u30fc\u30bf\u306f\u30e9\u30f3\u30c0\u30e0\u306b\u767a\u751f\u3059\u308b\u3082\u306e\u3068\u306f\u9650\u308a\u307e\u305b\u3093\u3002<\/p>\n\n\n\n

\n

\u5f93\u6765\u306e\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u3068\u7570\u306a\u308a\u3001AI\u306f\u7e4a\u7d30\u3067\u5bb9\u6613\u306b\u54c1\u8cea\u304c\u52a3\u5316\u3057\u3066\u3057\u307e\u3044\u307e\u3059\u3002\u4eba\u624b\u306b\u3088\u308b\u30c7\u30d0\u30c3\u30b0\u306b\u306f\u3001\u591a\u304f\u306e\u6642\u9593\u3068\u52b4\u529b\u3092\u8981\u3057\u975e\u5e38\u306b\u56f0\u96e3\u3067\u3059\u3002\u200d
\u682a\u5f0f\u4f1a\u793eCitadel AI<\/strong><\/a>\u00a0\u3067\u306f\u3001\u958b\u767a\u304b\u3089\u904b\u7528\u306b\u81f3\u308b\u5168\u3066\u306eAI\u306e\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u30b9\u30c6\u30fc\u30b8\u306b\u304a\u3044\u3066\u3001AI\u306e\u7570\u5e38\u3092\u77ac\u6642\u306b\u81ea\u52d5\u691c\u77e5\u30fb\u9632\u5fa1\u3059\u308b\u30b7\u30b9\u30c6\u30e0\u3092\u63d0\u4f9b\u3057\u3066\u3044\u307e\u3059\u3002
\u30d3\u30b8\u30cd\u30b9\u3084\u30b3\u30f3\u30d7\u30e9\u30a4\u30a2\u30f3\u30b9\u30fb\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u4e0a\u306eAI\u56fa\u6709\u306e\u65b0\u305f\u306a\u30ea\u30b9\u30af\u306b\u5bfe\u3057\u3001\u4f01\u696d\u306e\u7686\u69d8\u304c\u30d7\u30ed\u30a2\u30af\u30c6\u30a3\u30d6\u306b\u5bfe\u51e6\u9802\u3051\u308b\u3088\u3046\u3001\u5c11\u3057\u3067\u3082\u7686\u69d8\u306e\u304a\u5f79\u306b\u7acb\u3066\u308b\u3053\u3068\u3092\u9858\u3063\u3066\u3044\u307e\u3059\u3002
\u3054\u610f\u898b\u30fb\u3054\u8981\u671b\u306f\u3001info@citadel-ai.com<\/a>\u00a0<\/strong>\u307e\u3067\u304a\u77e5\u3089\u305b\u304f\u3060\u3055\u3044\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"

\u306f\u3058\u3081\u306b \u5f0a\u793e\u3067\u306f\u7c73\u56fd\u56fd\u7acb\u6a19\u6e96\u6280\u8853\u7814\u7a76\u6240(NIST)\u306e\u8a31\u53ef\u3092\u5f97\u3066\u30012022\u5e743\u6708\u306b\u767a\u8868\u3055\u308c\u305f\u5831\u544a\u66f8 \u201cNIST Special Publication 1270: Towards a Standard for […]<\/p>\n","protected":false},"author":1,"featured_media":1200,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[9],"tags":[],"class_list":["post-1090","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog"],"_links":{"self":[{"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/posts\/1090","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/comments?post=1090"}],"version-history":[{"count":0,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/posts\/1090\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/media\/1200"}],"wp:attachment":[{"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/media?parent=1090"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/categories?post=1090"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/tags?post=1090"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}