After years of being a nascent technology, artificial intelligence is becoming the real deal with services providers, according to AT&T's Chris Volinsky.
Artificial intelligence (AI) has quickly gone from science fiction to an everyday, and relied upon, element of how telecommunications companies do business. That isn't to say that it's not exciting: To have a platform that can outwit and overwhelm the experts is a potent and hugely useful tool.
Big telecommunications companies, with their huge networks and millions of customers, are especially well positioned to benefit from AI, according to Volinsky, the assistant vice president of big data research for AT&T Labs.
In Part I of a conversation with Telco Transformation, Volinsky suggested that customer care, security and boiling through hundreds of hours of video to assess the health of cell towers are great use cases for AI.
Telco Transformation: How has the use of the term "artificial intelligence" changed as time has passed and it becomes more deeply used by business?
Chris Volinsky: It's being used now as a marketing phrase, as opposed to a technical phrase, so it does incorporate a lot of things. Traditionally, I think AI was a term that was used to identify models, or programs, that could do things that made it seem like it was acting in a human way, things that we traditionally viewed as human, things that only a human could do, like facial recognition or voice recognition.
Now, I think it's a much broader term that encompasses anything having to do with automation, and automating things that humans traditionally have done. I think, for purposes of the examples here, that I'm going to talk about use cases. That might be a better way to think about AI.
TT: Is AI especially useful for telecom?
CV: I think it's definitely true that the scale of the network, the amount of data that we have available to us, traditionally has forced us to be data-driven and to develop techniques to analyze and act on large-scale data problems in the way that AI does.
Now, fast forward 20 years, and we've got these web-scale companies that also collect data on a similar scale, so I don't want to say we're the only ones who do it, but because of [AT&T's] history and our pedigree, it's in our DNA. And as AT&T migrates into other businesses, we're not just a telecom company anymore.
TT: What's really exciting to you in AI right now?
CV: One is a project we call "Predictive Care." This is for customers of our broadband network and TV service U-verse. We kicked off a project about a year ago in which we try and predict the customers who are going to call in with a customer care issue.
Traditionally, in customer care, a customer has a problem, they call you and they tell you about it and then you try and fix it as best you can.
We wanted to take a predictive approach and try and use all the data that we have to try and predict who might be calling us next week with a customer care issue. How do you know that? Well, we have a lot of data that we collect on the services used by our customers. In particular, for broadband service, you get a lot of telemetry from the network that indicates that maybe there's a certain amount of speed degradation, or packet loss, or some way that you can measure that the customer's having a sub-optimal experience.
You can do a supervised learning problem to show that if the data for a customer looks like this, or is trending a certain way, then they're likely to call you with an issue within the next week or two weeks.
We build that model in order to detect which customers had a high probability of calling us with a problem. As it turns out, when you study the data about what the resolutions to these problems are, often the first step that the rep takes is that they ask the customer to reboot their modem. What we realized is that we can reboot modems remotely.
What we saw was that this resulted in a reduction of customer care calls for a cohort of about 37 percent over the next month, and it resulted in a reduction in actual dispatches to those customer sites of I think about 30 percent.
TT: My understanding is that one form of AI is pure number crunching and another type of AI focuses on more human-like thinking. Which is AT&T using?
CV: Right. It's more cognitive modeling, trying to actually learn in a way that a human does. These are programs that are written by humans, so it always comes down to raw data manipulation and computer science. The phrase "deep learning" refers to a class of models that are built from what are called neural networks, that have a lot of complexity to them, and the result of their number crunching, I think, makes it seem like they have results from some kind of actual conscious thought from a computer.
But we know that's not true. It's just because the learning process is so complex that we don't even really understand how it got the answers that it does. In my group we don't do any explicit modeling of things like cognitive thought. What we do is try and model how humans solve problems.
For instance, have dozens of security analysts who notice an anomaly on the network and then collect dozens of data sources, and work on these big screens and try and coordinate and correlate the alerts, and try and understand what the root cause is. We can look at how they address a problem, and how they process one of these anomalies, and try and build a program that quote unquote "thinks the same way they do," so that we can automate the analysts' experience. But ultimately it's always data crunching.
I know there are AI practitioners who have a much more "science fiction-y" view of this, that the computers are really thinking, but to me it always comes down to the data.
TT: Is there another interesting another interesting use case that comes to mind?
CV: Yes. It's not that exciting maybe, but it's exciting from our perspective and the company's. AT&T has made a huge investment and a push in software-defined networks, and moving our network from a hardware-based network to a software-based network. Part of being a software-based network allows us to easily centralize and analyze the data that comes off of the network.
What we're doing inside the software-defined network is building a lot of what we call closed loops. Traditionally, again, in a regular network, there would be a big network operations center where there would be big screens of KPIs for [assessing] how the network is doing.
Now, in a software defined network, we're trying to build in these self-healing closed loops, so that a machine learning algorithm can determine if something's gone wrong and take the action to fix it without human intervention. When it provides that fix it can analyze how that fix worked. It feeds that information back into the algorithm, and the algorithm is constantly getting better over time, because it's getting this feedback about what works and what doesn't.
What may be a little bit more interesting to readers would be work that we’re doing with drones. We're not actually deploying this yet, but we have a group that is looking into flying drones up to cell towers to inspect the top of the cell tower, where the antennas are connected.
The drones are able to do video inspection of the cell towers. What we're trying to build is the capability to automatically detect through the video if there's something wrong. The drone might go up and take an hour of HD video that [now] a technician on the ground has to look through an hour of video to see if there's a problem.
If we can automatically detect the problem, then we can summarize the hour-long video into the three minutes that the technician needs to look at. They would be able to really drill down into it. We want to use the advanced modeling technology to save the time of the technician, so that instead of doing this mundane task of looking through an hour of video, they can focus on the stuff that they actually have to explore and take action on.
— Carl Weinschenk, Contributing Writer, Telco Transformation