Public Interest Research in Data Management and Machine Learning with Associate Professor Bill Howe, University of Washington

With over 80% of the planet impacted by urbanization, metropolitan areas have become loci of influence socially, economically, and politically.  As cities grow in size and complexity, the broad use of automated decision systems appears inevitable; as some argue, AI is the new electricity.  But the data that power these systems are deeply flawed — incomplete, incorrect, inaccessible, and systematically biased. As a result, decision systems trained naively on opportunistic data sources will, at best, calcify the status quo and at worst operationalize and legitimize the wrong answers.

In this talk, I’ll describe our work to address these issues through a multi-institutional, multi-disciplinary, and use-inspired research agenda in responsible urban analytics.  I’ll also describe a specific line of technical work on the EquiTensors project, where we are facilitating access to heterogeneous collections of urban datasets while adjusting for fairness and privacy.  As building blocks, we have developed algorithms for repairing data using causal models, incorporating fairness into spatio-temporal models, learning joint embeddings for multi-modal datasets (i.e., images and text), and applying differential privacy to large-scale, complex datasets.

I’ll end with some thoughts on the importance of city-university collaborations in advancing basic research, training students, and engaging on critical societal issues.