Feds Look to AI Solutions to Solve Problems from Customer Service to Cyber Defense
From Amazon’s Alexa speech recognition technology to Facebook’s uncanny ability to recognize our faces in photos and the coming wave of self-driving cars, artificial intelligence (AI) and machine learning (ML) are changing the way we look at the world – and how it looks at us.
Nascent efforts to embrace natural language processing to power AI chatbots on government websites and call centers are among the leading short-term AI applications in the government space. But AI also has potential application in virtually every government sector, from health and safety research to transportation safety, agriculture and weather prediction and cyber defense.
The ideas behind artificial intelligence are not new. Indeed, the U.S. Postal Service has used machine vision to automatically read and route hand-written envelopes for nearly 20 years. What’s different today is that the plunging price of data storage and the increasing speed and scalability of computing power using cloud services from Amazon Web Services (AWS) and Microsoft Azure, among others, are converging with new software to make AI solutions easier and less costly to execute than ever before.
Justin Herman, emerging citizen technology lead at the General Services Administration’s Technology Transformation Service, is a sort of AI evangelist for the agency. His job, he says: “GSA helps other agencies and prove AI is real.”
That means talking to feds, lawmakers and vendors to spread an understanding of how AI and machine learning can transform at least some parts of government.
“What are agencies actually doing and thinking about?” he asked at the recent Advanced Technology Academic Research Center’s Artificial Intelligence Applications for Government Summit. “You’ve got to ignore the hype and bring it down to a level that’s actionable…. We want to talk about the use cases, the problems, where we think the data sets are. But we’re not prescribing the solutions.”
GSA set up an “Emerging Citizen Technology Atlas” this fall, essentially an online portal for AI government applications, and established an AI user group that holds its first meeting Dec. 13. And an AI Assistant Pilot program, which so far lists more than two dozen instances where agencies hope to employ AI, includes a range of aspirational projects including:
- Department of Health and Human Services: Develop responses for Amazon’s Alexa platform to help users quit smoking and answer questions about food safety
- Department of Housing and Urban Development: Automate or assist with customer service using existing site content
- National Forest Service: Provide alerts, notices and information about campgrounds, trails and recreation areas
- Federal Student Aid: Automate responses to queries on social media about applying for and receiving aid
- Defense Logistics Agency: Help businesses answer frequently asked questions, access requests for quotes and identify commercial and government entity (CAGE) codes
Separately, NASA used the Amazon Lex platform to train its “Rov-E” robotic ambassador to follow voice commands and answer students’ questions about Mars, a novel AI application for outreach. And chatbots – rare just two years ago – now are ubiquitous on websites, Facebook and other social media.
In all, there are now more than 100,000 chatbots on Facebook Messenger. Chatbots are common features but customer service chatbots are the most basic of applications.
“The challenge for government, as is always the case with new technology, is finding the right applications for use and breaking down the walls of security or privacy concerns that might block the way forward,” says Michael G. Rozendaal, vice president for health analytics at General Dynamics Information Technology Health and Civilian Solutions Division. “For now, figuring out how to really make AI practical for enhanced customer experience and enriched data, and with a clear return on investment, is going to take thoughtful consideration and a certain amount of trial and error.”
But as with cloud in years past, progress can be rapid. “There comes a tipping point where challenges and concerns fade and the floodgates open to take advantage of a new technology,” Rozendaal says. AI can follow the same path. “Over the coming year, the speed of those successes and lessons learned will push AI to that tipping point.”
That view is shared by Hila Mehr, a fellow at the Ash Center for Democratic Governance and Innovation at Harvard University’s Kennedy School of Government and a member of IBM’s Market Development and Insight strategy team. “Al becomes powerful with machine learning, where the computer learns from supervised training and inputs over time to improve responses,” she wrote in Artificial Intelligence for Citizen Services and Government an Ash Center white paper published in August.
In addition to chatbots, she sees translation services and facial recognition and other kinds of image identification as perfectly suited applications where “AI can reduce administrative burdens, help resolve resource allocation problems and take on significantly complex tasks.”
Open government – the act of making government data broadly available for new and innovative uses – is another promise. As Herman notes, challenging his fellow feds: “Your agencies are collecting voluminous amounts of data that are just sitting there, collecting dust. How can we make that actionable?”
Historically, most of that data wasn’t actionable. Paper forms and digital scans lack the structure and metadata to lend themselves to big data applications. But those days are rapidly fading. Electronic health records are turning the tide with medical data; website traffic data is helping government understand what citizens want when visiting, providing insights and feedback that can be used to improve the customer experience.
And that’s just the beginning. According to Fiaz Mohamed, head of solutions enablement for Intel’s AI Products Group, data volumes are growing exponentially. “By 2020, the average internet user will generate 1.5 GB of traffic per day; each self-driving car will generate 4,000 GB/day; connected planes will generate 40,000 GB/day,” he says.
At the same time, advances in hardware will enable faster and faster processing of that data, driving down the compute-intensive costs associated with AI number crunching. Facial recognition historically required extensive human training simply to teach the system the critical factors to look for, such as the distance between the eyes and the nose. “But now neural networks can take multiple samples of a photo of [an individual], and automatically detect what features are important,” he says. “The system actually learns what the key features are. Training yields the ability to infer.”
Intel, long known for its microprocessor technologies, is investing heavily in AI through internal development and external acquisitions. Intel bought machine-learning specialist Nervana in 2016 and programmable chip specialist Altera the year before. The combination is key to the company’s integrated AI strategy, Mohamed says. “What we are doing is building a full-stack solution for deploying AI at scale,” Mohamed says. “Building a proof-of-concept is one thing. But actually taking this technology and deploying it at the scale that a federal agency would want is a whole different thing.”
Many potential AI applications pose similar challenges.
FINRA, the Financial Industry Regulatory Authority, is among the government’s biggest users of AWS cloud services. Its market surveillance system captures and stores 75 billion financial records every day, then analyzes that data to detect fraud. “We process every day what Visa and Mastercard process in six months,” says Steve Randich, FINRA’s chief information officer in a presentation captured on video. “We stitch all this data together and run complex sophisticated surveillance queries against that data to look for suspicious activity.” The payoff: a 400 percent increase in performance.
Other uses include predictive fleet maintenance. IBM put its Watson AI engine to work last year in a proof-of-concept test of Watson’s ability to perform predictive maintenance for the U.S. Army’s 350 Stryker armored vehicles. In September, the Army’s Logistics Support Activity (LOGSA) signed a contract adding Watson’s cognitive services to other cloud services it gets from IBM.
“We’re moving beyond infrastructure as-a-service and embracing both platform and software as-a service,” said LOGSA Commander Col. John D. Kuenzli. He said Watson holds the potential to “truly enable LOGSA to deliver cutting-edge business intelligence and tools to give the Army unprecedented logistics support.”
AI applications share a few things in common. They use large data sets to gain an understanding of a problem and advanced computing to learn through experience. Many applications share a basic construct even if the objectives are different. Identifying military vehicles in satellite images is not unlike identifying tumors in mammograms or finding illegal contraband in x-ray images of carry-on baggage. The specifics of the challenge are different, but the fundamentals are the same. Ultimately, machines will be able to do that more accurately – and faster – than people, freeing humans to do higher-level work.
“The same type of neural network can be applied to different domains so long as the function is similar,” Mohamed says. So a system built to detect tumors for medical purposes could be adapted and trained instead to detect pedestrians in a self-driving automotive application.
Neural net processors will help because they are simply more efficient at this kind of computation than conventional central processing units. Initially these processors will reside in data centers or the cloud, but Intel already has plans to scale the technology to meet the low-power requirements of edge applications that might support remote, mobile users, such as in military or border patrol applications.