Critical Data & AI Lecture Series #5. From Types to Tasks: Error and the Emergence of a Machine Learning Society

Lecture by Alex Campolo, Durham University.

Abstract

This talk compares statistical laws of error with a concept of error that has emerged in contemporary machine learning to draw out their distinctive epistemological and ethical implications. The first sense crystallized in the nineteenth century as practical techniques for producing estimates from discrepant observations were interpreted as metaphysical laws of error. Historians have shown how these interpretations, in turn, produced new forms of social knowledge of and control over normal types: races, nations, and classes. Although the metaphysical understanding of these statistical error laws was abandoned, error continued to be a major theme in twentieth-century statistical thought. In the middle part of the century, John von Neumann and Frank Rosenblatt proposed statistical paradigms of computing to allow machines to learn from their environments. In the 1980s, researchers developed techniques like backpropagation that used error measurements to improve a model’s performance on these learning tasks. Comparing these conceptions of error, I argue that we can perceive a shift from a politics of normal types revealed by the statistical regularity of error to a politics of tasks in which errors are used to refine desired behaviors in machine learning.

Bio

Alexander Campolo is a postdoctoral researcher on the Algorithmic Societies project in the Department of Geography at Durham University. He works on topics in the history and epistemology of machine learning and other data technologies. He has a special interest in situating machine learning within the history of statistics and probability and their associated governing rationalities. He received his PhD in Media, Culture, and Communication from New York University and has previously worked at the AI Now Institute and the Institute on the Formation of Knowledge at the University of Chicago. 

About the lecture series

The Critical Data and AI Lecture series is organized by Louis Ravn and Nanna Bonde Thylstrup as a joint venture between the research projects AI REUSE (DFF) and Data Loss: The Politics of Disappearance, Destruction and Dispossession in Digital Societies (DALOSS, ERC Stg).