Interpretability, or Learning to Listen to Algorithms

How do algorithms work? As algorithmic systems—from Google’s search engine to Spotify’s recommender—have become objects of popular concern, this question has proven vexing. Not only are these black boxes hidden from public view and illegible to the untrained eye, they are also complex, distributed systems. With the advent of techniques like deep learning, algorithmic systems are often described as “uninterpretable”—so complex that it is impossible, even for insider experts, to explain their outputs. And yet, engineers, like ordinary users, are tenacious interpreters, eager to make sense of algorithmic behavior, regardless of its internal complexity. In this talk, I draw on ethnographic fieldwork with developers of algorithmic music recommenders in the US to theorize “interpretability”, describing how engineers interpret supposedly uninterpretable systems. Music and listening offer useful models for making sense of this interpretive work, for the engineers as well as outside critics. Developers are not uniquely able to “see” inside algorithmic black boxes but rather learn to listen to them, and their own musical sensibilities are knit into the supposedly rational and quantitative operations of algorithmic systems.

Nick Seaver est professeur assistant au département d'anthropologie et au programme sur la Science, la technologie et la société de la Tufts University à Medford, dans le Massachusetts. Il étudie la manière dont les technologues donnent un sens aux préoccupations culturelles, y compris le goût et l'attention. Son prochain livre est issu d'une recherche ethnographique à long terme menée auprès de développeurs de systèmes de recommandation musicale aux États-Unis.

Entrée libre et sans inscription
Coordination : Nicolas Baya-Laffite et Boris Beaude