Smart machines at work: A.I. gets a job

Artificial intelligence and digital identity
Thinkstock

Second -- and third, and fourth -- opinions concerning cancer treatment. Analysis of the impact of social media publicity campaigns within minutes of an event. Analysis of the finances of a doctor's office written entirely by a machine.

These are a few examples of the new world of smart machines -- the latest generation of artificial intelligence (A.I.) configured to perform real-world tasks. The power of these machines is such that some pundits are predicting that they will contribute to displacement of workers and economic disruption. Those with hands-on knowledge, however, say these machines are still a long way from true human cognition.

Long road

"People have been trying to make smart machines for well over 60 years," notes Elliot Turner, head of AlchemyAPI, which sells an A.I. machine learning platform. "They started with neural networking, then developed expert systems, and then machine learning with shallow [manually crafted] learning. Those approaches underlie a lot of fraud prevention algorithms today. But the mid-2000s saw a revolution called 'deep learning.' They took neural networks and stacked them in a way that let computers learn very complicated representations of data."

Deep learning, meanwhile, has been made practical by the availability of massive amounts of computing power -- especially in the cloud -- and the availability of huge amounts of data on the Internet for training purposes, Turner adds.

Advances in natural-language processing have also been important enablers, notes Kenneth Brant, an analyst at Gartner. "These are cloud-scale applications; we are not talking about something that sits on a desktop," he cautions.

"In the next five years I expect that the range of decisions we let machines make autonomously will grow substantially," says Tim Estes, CEO and founder of Digital Reasoning, which offers a natural language processing engine.

"Over the next five or 10 years we will see the rise of a class of useful, personalized, A.I.-based machines that can do things for you," agrees Neff Hudson, vice president at insurer USAA. "It will proceed in stages, and now it's in the personalized scripting phase," in other words, the ability to respond to questions.

Helping in oncology

At the Memorial Sloan Kettering Cancer Center, such a system has been developed to help doctors decide on the initial treatment for lung cancer patients. Currently being field-tested, it doesn't issue opinions, but rather lists possible answers with estimates of the confidence it has in each answer, explains Dr. Mark G. Kris, lead physician on the project.

Called Watson Oncology because it is based on the IBM Watson A.I. platform, it arrives at its decisions by analyzing multiple forms of medical literature, research reports and patient medical records, Kris notes.

"A lot of time was spent getting it to read medical records," explains Kris. "And while it's good at reading text from a textbook, reading tables from journal articles was something it had to be taught, as information in journal articles is often embodied in tables."

Indeed, doctors seeking treatment information often have to extrapolate it from reports of clinical trials since there is no specific medical literature on a large number of treatments, he explains. Consequently, Watson Oncology has to ingest information concerning thousands of ongoing trials.

"It frees the doctor to be a doctor, and to look the patient in the eye and say, 'I had a machine distill the information; here's what it said. Now let's make a decision.' The patients really like the idea of an independent party looking at the work," Kris says.

With the oncology field's nonstop developments, the system will need constant upgrades, and will never be finished, he says. The system will also track the outcome of its own patients. "The prize will be to show, through analytics, that the decisions that were made led to good outcomes, and those outcomes can be the basis for decisions on similar patients," Kris says.

For customer service

At USAA, IBM Watson was one of the platforms being used to create what Hudson calls question-answer machines, intended for customer support.

With such machines, "The user does not just issue a command and have it obeyed, but asks it a question and gets a nuanced answer," Hudson notes. The USAA smart machines include a virtual agent that can discuss about 200 different topics, including what to do after a car accident, he says. USAA has also launched a machine that offers advice to members who are about to leave the military, and is working on one that would advise younger members on personal finances. Hudson says he's contemplating giving that one a drill sergeant voice.

He reports numerous challenges. For instance, "We have found that people like to switch rapidly between general and personal topics, from, 'What's the best car?' to 'How can I afford it?' Computers are not good at that yet," he says.

Meanwhile, natural language interfaces must take into account the preference of older users to issue discrete commands, while younger ones prefer talking naturally to a computer, he indicates. In either case if the system doesn't answer within five seconds, the users will pose the question again, burdening the system with another thread, he adds.

For reports and PR

A smart machine built for Zotec Partners, a medical billing service, uses a natural language interface, but in this case, it's for output, not input. Jeff Maze, Zotec's director of business intelligence, says the idea was to automatically generate monthly two-page financial reports for each of the several hundred medical groups that are the firm's clients.

"The head of the group should be able to read the two pages and have a clear understanding of the group's financial performance and any trends that should be noted," says Maze.

The machine is based on an off-the-shelf product called Quill from Narrative Science that generates narrative reports from feeds of the kinds of structured data that enterprises usually have on file, such as transaction data and performance data, explains Stuart Frankel, head of Narrative Science.

"We've gotten some very encouraging feedback from clients," Maze adds, but he also notes that the machine's development wasn't sparked by customer requests. Instead, it was inspired by Maze's experience with fantasy football, where a witty summary was generated at the end of each game.

Immediate feedback was also the goal of public relations firm Waggener Edstrom Communications, which built a machine called Waggener Edstrom Infinity (WEI) to gauge, in near real-time, the impact of a client's message (or latest publicity event). The firm used the AlchemyAPI platform, which combines deep learning with unstructured data.

"Our clients are faced with a deluge of data, but now they have a way to capture the right data to make the right decisions," says Karla Wachter, senior vice president at Waggener Edstrom. "We could not find what we wanted in the marketplace, so we built it in-house."

Filtering news and social media feeds for mentions of the client, the system gauges impact by measuring perception (i.e., the sentiments being expressed), amplification (i.e., whether people are sharing the message), engagement (whether people are posting comments about it) and reach (how many people saw the message), Wachter explains.

"The level of reporting is richer and more strategic now, and clients can understand what messages are landing and what messages are not," she says.

"We continue to work on it, adding more data feeds or making it more accessible. We will never be finished," adds David Kohn, Waggener Edstrom's vice president of development.

Scale of effort

What the above examples have in common is the nontrivial effort involved in creating a smart machine. USAA, for instance, has a core group of 20 employees working on the firm's smart machine projects, says Hudson.

The Watson Oncology system took three years to develop, with 10 people from Memorial Sloan Kettering and another 10 from IBM, notes Kris.

At Waggener Edstrom, Kohn says an in-house team of 10 worked on the solution for three years.

The Zotec system took a full year to create, explains Maze, with two people from Zotec and another from Narrative Science. "It was like hiring a new person who hasn't worked in healthcare before and giving them a crash course in healthcare, billing and Zotec, and making them independent enough to review and analyze a client, and draw out trends," Maze recalls.

Looking forward, building smart machines "will mature to a cottage industry, after we get to the point where we understand the technology so well that we can see common patterns and simplify what is available," says Rob High, CTO of IBM's Watson Group, vendor of the Watson A.I. platform. He notes that Watson-based projects can currently take anywhere from six weeks to 18 months to deploy, but his group is working on a tool that would permit even more rapid deployment of an "engagement adviser" that would answer questions about a product or service, High adds.

1 2 Page 1
Page 1 of 2
Bing’s AI chatbot came to work for me. I had to fire it.
Shop Tech Products at Amazon