National Frontiers panelists discuss how artificial intelligence can affect policy
On Thursday, before President Obama had arrived, the Frontiers conference began with several sessions explaining how artificial intelligence (AI) and data science affect various levels of policy. The local and national sessions, called tracks, took place on the second floor of the Cohon Center. The panels were a who’s who of academic and corporate giants in the field of AI such as Engineering Director at Uber’s Advanced Technologies Center Raffi Krikorian and Corporate Vice President of Microsoft Research Jeanette Wing. The local policy track also featured municipal government officials who embody the spirit of making data “work for us” as the president often says.
The National Frontiers track opened with Pennsylvania Governor Tom Wolf telling the story of innovation in Pennsylvania, beginning with William Penn’s idea of freedom of conscience, winding through the founding ideas of American government born in Pennsylvania, touching on the agricultural and industrial revolutions, and finally landing on the transportation revolution which still roils on today as self driving cars become frequent sights on the streets of Pittsburgh. “Welcome to Pittsburgh; welcome to Pennsylvania; welcome to the future,” said Wolf before he ceded the stage to a series of “lightning speakers” who gave ten minute talks on their areas of interest or expertise.
The first speaker was Tanya Berger-Wolf of the University of Illinois at Chicago. She spoke about how data science is making it possible to track animals without invasive tracking devices that could be dangerous to the animals. The program she referred to was able to use machine learning techniques to identify species and even individual animals. This resulted in programs like Flukebook for tracking whale flukes and the IoT, Internet of Turtles.
The second speaker was Carnegie Mellon robotics professor Stephen Smith. Smith presented a plan to alleviate urban congestion by having traffic lights autonomously optimize their own patterns and then send that information downstream so future lights know what to expect. This program found great success in Penn Circle, reducing travel time by 26 percent, idling time by 41 percent, and emissions by 21 percent. After a successful expansion to Bakery Square, Smith’s team plans to continue expanding the program throughout Pittsburgh.
The final lightning speaker, Suchi Suria of Johns Hopkins, explained how electronic health records could be used to diagnose conditions such as sepsis that require a quick response. The algorithm she and her team developed was able to use these records to discover sepsis up to 24 hours before doctors are able to make the same diagnosis. This approach was not unique to sepsis, and has also been successful in responding to other conditions and could help diagnostics significantly.
Krikorian and Wing, alongside founding director of the New York University Center for Data Science and Director of AI Research at Facebook Yann LeCun, Vice President of Cognitive Computing at IBM Guru Banavar, and research scientist at Google Francoise Beaufays, participated in a panel on best practices in AI. The panel began with how AI developed from something of a niche field of study into the dominating presence it has today. All of the members of the panel agreed that the convergence of the available tools — humans now having powerful computers in our pockets — and deep learning in the last five years have expanded the capabilities of AI to the point where they are unavoidable. Deep learning is a process by which an algorithm attempts to model abstractions from data, allowing for more general decisions. Wing explained that AI began as a dream of recreating human intelligence, but that quickly proved challenging. Eventually, people identified individual tasks that a human can do and attempted to algorithmically recreate them which has lead to the specialized AI we have today.
The panel also got into some of the specific limitations that computers face. One specific example that LeCun gave was the difference between the sentences “the trophy did not fit in the suitcase because it was too large” and “the trophy did not fit in the suitcase because it was too small.” Humans observe the world and are easily able to discern that either the suitcase was too small to contain the trophy or the trophy was too large to fit in the suitcase, but that observation of reality is missing from computers. Wing added that conversational English is different from written English and sometimes people will switch languages when they are speaking. LeCun believes that these challenges have to be overcome before AI can start to function as accurately as humans.
Finally, Wing mentioned the importance of secure AI that is safe from some of the more harmful tendencies of people. This was in response to a question about Microsoft’s Tay, which has been described as a “Hitler-loving sex robot.” Tay was originally supposed to be an automated conversational bot that picked up on people’s tendencies, but was quickly corrupted into saying some things that were less than friendly. This came on the heels of a wildly successful version of Tay named XiaoIce which was released in China. It showed the need to actively make sure AI doesn’t become harmful.
In the Local Frontiers track, Chief Data Scientist at the University of Chicago Center for Data Science and Public Policy Rayid Ghani, Knoxville Chief of Police David Rausch, and Commissioner of the Massachusetts Department of Public Health Monica Bharel discussed how data can be used to disrupt cycles of incarceration. Rausch’s police department has gone two years without a police involved shooting, and when asked about how that happened, he responded that there was a little luck but a lot of effort. He mentioned that community engagement helped officers and their communities interact more positively. He also talked about how the education of officers, including a large increase in the number of officers with four year degrees, and more stable working conditions and hours for police officers helped to make police less likely to need violence. He also talked about how officers are trained to reset situations where they are at a disadvantage so they never have to resort to violence. Ghani added that these were all backed up by data saying these were meaningful predictors of violence. Police who have four year degrees are less likely to commit an act of police violence while police who consistently respond to distressing situations such as domestic violence or suicides are more likely to commit an act of violence in the coming weeks.
Bharel also added the importance of tackling substance abuse to disrupting cycles of incarceration and talked about how Massachusetts implemented the State Without Stigma program to help create vocabulary to talk about drug use and addiction in a way that doesn’t alienate people. She also spoke about how the data supports that the start of the opioid crisis is not heroin and illegal drugs, but that this comes as a result of opioid prescriptions for things like back pain. The data explains the flow of the problem and helps the department respond.
The tracks were heavy on specific policy discussions and how data science affects how both private and public entities should respond to the vast amount of information that now surrounds us. They were followed by the plenary session featuring President Obama that focused on brain science and medical information.