On the genealogy of machine learning datasets: A critical history of ImageNet

Abstract

In response to growing concerns of bias, discrimination, and unfairness perpetuated by algorithmic systems, the datasets used to train and evaluate machine learning models have come under increased scrutiny. Many of these examinations have focused on the contents of machine learning datasets, finding glaring underrepresentation of minoritized groups. In contrast, relatively little work has been done to examine the norms, values, and assumptions embedded in these datasets. In this work, we conceptualize machine learning datasets as a type of informational infrastructure, and motivate a genealogy as method in examining the histories and modes of constitution at play in their creation. We present a critical history of ImageNet as an exemplar, utilizing critical discourse analysis of major texts around ImageNet’s creation and impact. We find that assumptions around ImageNet and other large computer vision datasets more generally rely on three themes: the aggregation and accumulation of more data, the computational construction of meaning, and making certain types of data labor invisible. By tracing the discourses that surround this influential benchmark, we contribute to the ongoing development of the standards and norms around data development in machine learning and artificial intelligence research.

Reading datasets: Strategies for interpreting the politics of data signification

“…The methods presented in this article offer an alternative and potentially complementary approach to examining bias, contributing to efforts to localize (Loukissas, 2017), critique (Beaton, 2016), and contest (Denton et al., 2020) datasets. These methods draw on frameworks from the humanities rather than STEM fields for data critique, prompting us to treat datasets as cultural artifacts refracting the social and political contexts of their production as opposed to value-neutral artifacts that become distorted through special interest politics.…”

“…While there are intersecting aims for engaging each of these methods, a distinct aim of a connotative reading is to situate data semantics historically and culturally in order to interpret how implied meanings are derived from data. In this sense, connotative readings advance efforts to document a “genealogy of datasets” (Denton et al., 2020). Sometimes, information pertinent to a connotative reading is written up in thoughtful data documentation.…”

Artificial intelligence, algorithms, and social inequality: Sociological contributions to contemporary debates

“…Sociologists can theorize these developments by examining how social inequalities are structured, highlighting political economy, capitalism, and colonial relations (Couldry & Mejias, 2019; Dyer‐Witheford et al., 2019; Shestakofsky, 2020). While macro‐level social theories provide analytic tools for global transformations, sociologists can attend to the production of power and knowledge through genealogies (Denton et al., 2021) and ethnographies of AI research (Hoffman, 2021; Jaton, 2021). There will also be continuing value in producing ethnographies (and institutional ethnographies, James & Whelan, 2021) of organizations implementing algorithmic systems (Bailey et al., 2020; Brayne & Christin, 2021; Cruz, 2020; Shestakofsky & Kelkar, 2020), as well as studies into the experiences of people who are further ‘downstream’, interacting with algorithmic systems (Christin, 2020; Noble, 2018).…”

Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content?

“…However, even ImageNet [9], which was released in 2012 and remains one of the most popular datasets in the computer vision domain to this day [4,46], contains questionable content [3]. The entailed issues have been discussed for language models, for instance, models producing stereotypical and derogatory content [2], and for vision model respectively CV datasets highlighting, e.g., gender and racial biases [10,29,44,48].…”