Garbage In, Garbage Out: The Predictable and Unpredictable Challenges of Regulating Machine Learning Systems

Share

To attend remotely, register and request Webex information.


We are in a new golden age of artificial intelligence research, eclipsing the postwar zenith and driven by a fundamentally different conceptual approach. It's powered by big data, new storage and processing capacities, and pattern-discrimination systems that learn from those data, constructing rules for their behavior as they go. This isn’t just in the lab. These machine learning systems are implemented by all the big tech companies in everything from ad auctions to photo-tagging, and are supplementing or replacing human decision making in a host of more mundane, but possibly more consequential, areas like loans, bail, policing, and hiring. And we’ve already seen plenty of dangerous failures: Computer vision systems that don’t recognize black faces or classify them as gorillas, risk assessment tools systematically rating black arrestees as riskier than white ones, hiring algorithms that learned to reject women. These issues force a fundamental reconsideration of core democratic values—not just in what decisions are made, but how they are reached, and with sort of accountability. This talk will review fundamental issues of fairness and equity in machine learning systems and demonstrate how they play out in the specific domain of policing. Finally, we will discuss emergent approaches for designing, auditing, and regulating these systems, and what we can learn both from other fields that have faced similar conflicts, and from activists on the ground. 



  Date and Time

  Location

  Contact

  Registration



  • 4450 Wisconsin Ave NW
  • Metro Station: Tenleytown-AU
  • Washington, District of Columbia
  • United States
  • Building: Tenley-Friendship Neighborhood Library
  • Room Number: Large Meeting Room (2nd Floor)
  • Click here for Map
  • Murty Polavarapu

  • Starts 22 June 2019 02:59 PM
  • Ends 25 September 2019 05:59 PM
  • All times are America/New_York
  • No Admission Charge
  • Register


  Speakers

Dr Daniel Greene of University of Maryland, College Park, MD

Topic:

Garbage In, Garbage Out: The Predictable and Unpredictable Challenges of Regulating Machine Learning Systems

We are in a new golden age of artificial intelligence research, eclipsing the postwar zenith and driven by a fundamentally different conceptual approach. It's powered by big data, new storage and processing capacities, and pattern-discrimination systems that learn from those data, constructing rules for their behavior as they go. This isn’t just in the lab. These machine learning systems are implemented by all the big tech companies in everything from ad auctions to photo-tagging, and are supplementing or replacing human decision making in a host of more mundane, but possibly more consequential, areas like loans, bail, policing, and hiring. And we’ve already seen plenty of dangerous failures: Computer vision systems that don’t recognize black faces or classify them as gorillas, risk assessment tools systematically rating black arrestees as riskier than white ones, hiring algorithms that learned to reject women. These issues force a fundamental reconsideration of core democratic values—not just in what decisions are made, but how they are reached, and with sort of accountability. This talk will review fundamental issues of fairness and equity in machine learning systems and demonstrate how they play out in the specific domain of policing. Finally, we will discuss emergent approaches for designing, auditing, and regulating these systems, and what we can learn both from other fields that have faced similar conflicts, and from activists on the ground. 

Biography:

Daniel Greene is an Assistant Professor of Information Studies at the University of Maryland College Park. Prior to joining UMD, he was with Microsoft Research New England. Daniel's research focuses on the future of work, using sociological approaches to explore the fights to define that future in workplaces, training institutions, and technological designs. He is currently at work on a book manuscript for the MIT Press that draws on years of ethnographic research to investigate the impact of learn-to-code curricula on both students and the schools and libraries deploying them.  A separate but related stream of research explores the history, design, and impact of surveillance systems.  With computer vision researcher Genevieve Patterson, Daniel recently published “The Trouble With Trusting AI to Interpret Police Body-Cam Video” in IEEE Spectrum, and he’s excited to discuss these issues with practitioners. You can find him online at dmgreene.net.

Address:University of Maryland, , College Park, Maryland, United States





Agenda

6:30 PM to 7:00 PM - Refreshments and Networking

7:00 PM - 7:05 PM - Chapter announcements and Speaker Introduction

7:05 PM - Talk