核心概念
本文提出了一種基於協作設計和公平、多元化和包容性 (EDI) 原則的方法,用於創建一個評估公共空間品質的資料集和 AI 模型,重點關注代表性不足群體的觀點,並探討了在 AI 模型中捕捉多元化觀點所面臨的挑戰。
統計資料
資料集包含從大蒙特婁地區收集的 7,833 張街景圖像。
資料集包含 19,990 對圖像的成對比較。
共招募了 28 名參與者參加研討會和圖像標註。
招募工作針對代表不同代表性不足群體的各種社區組織。
在參與者中,20 人認為自己是女性,5 人屬於少數民族,2 人是殘疾人,10 人是 LGBTQ2+ 社區的成員,2 人屬於宗教少數群體。
為了捕捉公共空間的多樣化用途,我們定義了 35 個評估其品質的標準。
引述
"Current advancements in AI heavily rely on the availability of large-scale datasets meticulously curated and annotated for training purposes."
"However, concerns persist regarding the transparency and context of data collection methodologies, particularly in instances where annotations are sourced through crowdsourcing platforms."
"To address these limitations, we propose a methodology grounded in a specific socio-cultural context for dataset collection and AI model development."
"Our approach centers on a co-design model that actively involves stakeholders at key stages of the AI model development, including dataset creation."
"Additionally, we integrate principles of Equity, Diversity, and Inclusion (EDI) to ensure diverse viewpoints are represented within the dataset."