Federal agencies overwhelmed by data and struggling with increased workloads are increasingly considering artificial intelligence solutions but they won’t find a well-trodden path.
“There is more going on in this space than people know or appreciate throughout the federal government,” General Services Administration’s Emerging Technology Office lead Justin Herman told the audience at the ACT IAC Artificial Intelligence Forum last week.
Last year, the GSA office launched a working group dedicated to AI in agencies and is collecting agency use cases on Github so others can learn from the pioneers.
» Get the best federal technology news and ideas delivered right to your inbox. Sign up here.
GSA, for example, has a proof-of-concept that uses blockchain and automation to speed up the procurement process. The IRS is looking at options like intelligent voice agents and sentiment mining to improve service delivery and is experimenting with AI tools to cut down on fraud and identity theft in the more than 3 billion financial documents it receives per year. The National Institutes of Health’s National Institute on Aging uses IBM’s Watson to map genomes and detect anomalies for a moonshot program to find to a treatment for Alzheimer’s disease by 2020—and the project already has a petabyte of data with four more years of sequencing to go.
Getting management buy-in and budget for new technology will often be a challenge for federal agencies, but adopting AI has some additional challenges. Here are a few:
There’s an AI boom.
NVIDIA, a company that creates AI chips and platforms, worked with 1,500 companies focused on accelerating AI in 2014. The company now works with more than 39,000, as well as federal agencies including the Energy, Defense, Commerce departments and NASA, Senior Solutions Architect May Casterline said.
When IT consulting firm Riva Solutions started researching companies that offer customer service technology, it found pages and pages of providers instead of the handful it was expecting, Chief Technology Officer Raj DasGupta said.
Industry partners are out there, but might not work with the government yet, Herman said.
AI adoption won’t be plug and play.
When agencies first started adopting cloud computing, it required highly skilled set of IT professionals. As cloud advanced, companies created off-the-shelf kind of tools and services that don’t require that same expertise. For now, those don’t exist for many AI solutions, which means agencies will need to recruit data scientists.
“AI is totally ‘Mad Max’ right now,” Dent said. “It’s everyone on their own little cart driving around in the desert trying to figure out what direction things are going to go. There’s not productized framework that’s the solution for AI that everyone uses.”
Agencies need training data.
Quality training data is still a challenge for many projects and training systems is still a “manually intensive process,” Casterline said. For example, teaching an AI system to recognize a stapler requires many, many images of staplers with the word “stapler” on it.
AI still needs people—for now.
For now, most AI systems require a human in the loop to make the final decisions. “The technology isn’t ready to make decisions; it can make suggestions,” NVIDIA’s Casterline said.
This is in part because these systems are fragile and don’t adapt well to what to humans may seem like small changes, Jeff Alstott, Intelligence Advanced Research Projects Activity project manager said. Google’s AlphaGo system, for example, can beat human players at the game Go, but Alstott said changing the size of the board messes the system up.
Many systems that perform well in training environments have issues once deployed, he explained. Both IARPA and the Defense Advanced Research Projects Agency are actively researching ways to make these systems more robust
“We’re not ready to hand everything over to machines quite yet,” Alstott said.