Sponsored By
During the recent MassIntelligence Conference in Boston, tech pundits and industry watchers convened to talk about the big data item on every enterprise's virtual lips: artificial intelligence (AI).
"Seventy-six percent of enterprises are saying that AI is important," said Mike Frane, vice president of product management at Earthlink, the newly acquired division of telco Windstream Communications Inc. (Nasdaq: WIN) "They don't know how, but it is."
"Machine learning is eating the world," said Sam Madden, a professor of electrical engineering and computer science at MIT, in a keynote address. "You can use machine learning algorithms to optimize whatever TCP/IP environment you're running in better than a human could."
Madden went on to explain that, along with algorithms, the other two elements that artificial intelligence is primarily broken down into are those of "massive" compute power, often delivered via mega-datacenters -- think Amazon and Google -- to enable the algorithms, and the data itself.
It is these two elements with which network carriers, such as Windstream are especially concerned -- largely because they are directly tied to a fourth element that Madden did not directly address: connectivity.
Computational connectivity
There is a peculiar correlation between connectivity and compute power in the AI context: They both feed into each other.
"Windstream is not an AI company, but we see the trends that are driving AI and impacting customers' networks and the way that customers run their businesses," said Frane. "The demands of AI, the demands of cloud, are gonna require new infrastructure."
SD-WAN, for Frane and his colleagues, perhaps best represents this new, "smart" infrastructure solution for network optimization in the AI age because SD-WAN allows real-time network monitoring and optimized traffic redirection.
Indeed, Frane reported a "tremendous surge in the adoption of SDN" in general among enterprise customers to compensate for the inherent lack of intelligence at the network edge. Faced with this increased demand, Windstream launched its SD-WAN solution in the fourth quarter of 2016. (See Windstream Targets Mid-Market With SD-WAN.)
"As we've started to embrace SD-WAN, we have been able to dovetail into AI with some of our capabilities with the network," said Frane. "We see how important connectivity is. It's looking at AI from the connectivity perspective from the end device to the cloud, and then ensuring that that connectivity is available as close to 100% of the time as possible, and understanding that not all data is created equal."
From there, Frane explained that SD-WAN systems can -- and will have to -- appropriately hasten delivery of the data required for AI solutions while re-prioritizing less critical data packets along different network paths. In the age of cloud centralization, he continued, this kind of infrastructure is vital.
Frane's hope for the future is that the very SD-WANs that Windstream is deploying to enable AI solutions will one day themselves be artificially intelligent enough on their own, via machine learning, to anticipate traffic demand and requirements, and auto-schedule traffic redirection schemes based on time and day.
"Once you have the information, it's going to be critical that the devices you design, or integrate, have the ability to consume that information in real time," he said. "To deliver the promise of AI, the data has to be transferred instantly and all the time."
Data determinism"It's all about the data and what you can do with it," Robert Mawrey, IoT Principal of MathWorks, said in a separate panel session with Frane. "There are a couple of challenges, so it's awful hard to know what the context is."
In his keynote, Madden explained that AI systems are highly data reliant because they are not so much encyclopedias of knowledge as they are "systems of learning." Consequently AI -- at least to the extent it exists today -- is but a complement to actual human intelligence.
"How can we use computers to perform these human-like tasks to augment what humans do with this flood of data that we now have?" Madden rhetorically asked. "We have these algorithms just to do things that you never thought would be possible."
More to the point, when possibilities cannot be imagined, they cannot be predicted. Carl Kraenzel, vice president of security and support for Watson Cloud at IBM, further discussed these ideas during an onsite interview about IBM's Jeopardy-winning Watson after a breakout session in which he participated.
"If you tried to create the ontology for Jeopardy, you would always fail," Kraenzel said. "Rules engines are brittle when you go outside the boundary."
Whereas ontologies "are good for directing a workflow," Kraenzel explained, the better data-management solution for AI lies in "several annotators" all engaged in experiential machine learning because the unimagined "possibilities" to which Madden referred to by their definition go unpredicted.
Moreover, there are other, more practical problems of deploying AI systems that are more data-heavy.
"There's always a sort of 'chicken and egg' thing with data when you're a startup," said TechCrunch reporter Ron Miller, moderating Mawrey and Frane's panel. "How do you get popular enough to get enough data?"
Frane, for his part, saw hope for AI solutions at large and small enterprise customers alike because of the high availability of and high demand for data.
"With the explosion of capabilities in compute and storage and all the data we have in the cloud, and all these companies starting to open up their AI algorithms, I think it's really going to accelerate these smaller companies," said Frane, speaking on the same panel -- despite what he and other speakers referred to as a talent shortage in AI. "I think that once the large companies start to move into AI with some real vigor, it's going to generate a groundswell. I can't help but think that's going help the startups."
— Joe Stanganelli, Contributing Writer, Telco Transformation
