“Humans in the Loop: Humanities Hermeneutics and Machine Learning.” Keynote for DHd2020 (7th Annual Conference of the German Society for Digital Humanities), University of Paderborn, 6 March 2020.
- Abstract: As indicated by the exciting new field of “interpretability studies” in artificial intelligence research, contemporary machine learning and data science create fundamental problems of interpretation. These issues of “explainability” are related to changing, computationally-inflected, and often antithetical views of knowledge as both generalizable and domain-specific, abstract (a “model”) and experiential (a “ground truth”), supervised and unsupervised, and open and reproducible. Perhaps the least understood dimension of machine learning is the “human in the loop” problem: how humans can or should engage sociologically, ethnographically, politically, institutionally, ethically, and hermeneutically in the processes of machine learning. In philosophical terms, how does the “hermeneutical circle” affect the human in the loop?
In this talk, Alan Liu uses the example of the “interpretation protocol” for topic models he is developing for the Mellon Foundation funded WhatEvery1Says project (which is text-analyzing millions of newspaper articles mentioning the humanities) to reflect on how humanistic traditions of interpretation might contribute to machine learning. Reversing this direction of thought, he also reflects on how machine learning might change the understanding of humanistic interpretation itself through fresh ideas about the relation between wholes and parts, similarity and difference, and “representations” and “models.”