Trading Expertise in Machine Learning Markets

• In Event: Unpacking Expertise - II: Authority, Politics, & the State

Abstract

What happens to data science and engineering expertise as machine and deep learning algorithms commoditize? Per Marx, the "magic and necromancy" of commodities cajole two parties to embark on an exchange. Forgotten in the trade is the work already done and yet to come. Algorithms materialize data scientists’ claims to expertise and promise repeatable outcomes in the hands of others. Often, they do deliver. But the ease of the exchange obscures the data wrangling work necessary for them to do so.

In prior work, I characterized how computer vision, including deep learning, algorithms change hands as commodities. In this paper, I call on this and additional ethnographic research with now over eighty machine and deep learning professionals from engineers who deploy neural networks to count endangered fish to data entry managers who ensure shipping manifests describe what is actually onboard. I explore the resulting data and engineering work fallout from open source and public algorithmic marketplaces. These marketplaces effectively shift the burden of training data collection, annotation, management and compliance to engineers and their organizations. They are, frankly, ill prepared for this social and technical work and find themselves forging new work practices, many of which they might have prior deemed below their pay grade.

Edward Hutchins and Lucy Suchman's definitions of professional work undergird my argument that expertise, like epistemologies, are partial and situated. They also distinguish, per Pierre Bourdieu. And when traded, they shift the distribution of who does what algorithmic and data work.

Author

• Suzanne L. Thomas, Intel Labs