2021 Lectures

How 1950s ideas about creativity in machine learning continue to inform social and political possibility today

Mendon-Plasek, Aaron

Adelle and Erwin Tomash Fellowship Lecture in the History of Information Technology, Charles Babbage Institute, University of Minnesota, 5 May 2021.

In 1982 Newell observed retrospectively that by 1955 machine learning researchers had already splintered off from those who "became the AI community.” If early machine learning wasn’t artificial intelligence, what was it? This talk traces the historical problems of knowledge to which 1950s to 1970s "machine learning" was presented as a solution, and discusses why machine learning problem-framing strategies came to be seen by some as a compelling strategy for adjudicating social questions. Drawing on a study of thousands of published technical articles and original research in more than 12 archives, I illustrate how a collection of problem-framing strategies researchers denoted "machine learning" valorized a distinctive conception of creativity as the “production of completely new ideas not deducible from known data.” I show how efforts to instantiate this particular notion of creativity in digital computers was shaped by professional, social, and intellectual challenges of the researchers involved, and resulted in the creation of "learning machines" designed to (1) make human-like judgments given messy, unexpected, and even contradictory data, and that could (2) redefine “significance” for a given task apart from the programmer’s understanding of the task to be performed. Flipping the script of co-eval artificial intelligence research rooted in proof-proving and operations research, mid-century machine learning researchers argued that the lack of explicit definitions of relevant concepts or rules necessitated by learning directly from examples gave machine learning systems the capacity to respond to a greater range of contexts than those system builders could otherwise specify pragmatically. Emerging from the diffuse, transnational network of researchers attempting to implement creativity on digital computers in the 1950s and 1960s were two crises for individual researchers: (1) the practical inability to compare different learning machines that ostensibly did the same task, and (2) the existential inability to articulate what made machine learning qua pattern recognition a unique discipline. The solutions devised by machine learning researchers in the 1960s to address these two crises offered new modes of scientific identity for individual researchers and enlarged what constituted a legitimate description of the physical and social world. These descriptions of being and knowing continue to frame possibility in ongoing debates regarding the relationships between AI and society today.