Session Name: Active Networks: Thinking Beyond Just Advertising For Machine Learning.
Speakers: S1 Announcer S2 Jo Maitland S3 Jeffrey Davitz S4 David Gutelius
…who’s the CEO of Solariat, and David Gutelius the Chief Social Scientist at Jive Software about Active Networks: Thinking Beyond just Advertising for Machine Learning – and they’re already here, give them a round of applause [applause].
JO MAITLAND 00:15
So just briefly to set the stage for you guys, these two gentlemen that we have with us here this afternoon previously formerly worked for SRI. They were behind a DARPA project that essentially set up the biggest machine learning and AI that we’ve seen so far in the industry, to give you a little bit of context. We want to talk about the concept of active networks – what that actually means. So Jeffrey I’m going to kick that to you.
JEFFREY DAVITZ 00:43
A lot of the kinds of networks that people see, and there’s been some discussion of this about how networks are these fabulous sources – and I mean by social networks – of big data. But largely speaking these networks have been very passive – they don’t do much in the way of trying to drive certain kinds of behavior, other than through advertising. So it’s very peripheral to the function of the network which is mostly about people sharing with each other. But there’s a whole class of active networks which is growing now as people realize that social networks do lots of work – I mean social networks historically have always been the context within which lots of work gets done.
JEFFREY DAVITZ 01:26
What we mean by active social networks is using AI – generally speaking machine learning – in order to facilitate and motivate networks to accomplish certain kinds of goals. For example you see this with Wikipedia, you see this with the generation of Frequently Asked Questions, generating knowledge, making decisions, doing all kinds of things social networks have done. What you can do with the big data that you have now that you never had with social networks before is you can use it to intelligently guide and motivate the social network to accomplish certain kinds of goals. That’s what we mean by active social networks.
JO MAITLAND 02:05
So accomplish what kind of goals?
JEFFREY DAVITZ 02:07
Well, let me turn that over to Dave.
DAVID GUTELIUS 02:10
It runs the gamut from creating a more efficient more effective working groups inside of an enterprise. If you think of just in the news recently we’ve seen a lot about a new generation of cyber threats that are out there called advanced persistent threat actors. These are very hard groups to get a handle on and the general way that cyber is done today is at that individual defender level where you’ve got your proto-typical overweight guy in the dark in front of a Unix terminal watching the network traffic and trying to do something useful with it. What we’re trying to do is create – with this active networks concept – is create a way where machines are actually augmenting what humans are doing already but doing together. Creating a collective capability that is not new.
DAVID GUTELIUS 03:03
Jeffrey and I did a bit of this in the SRI context with this very large program called the KLO project – stands for Cognitive System that Learns and Organizes – very much a DARPA initiative. What we did was created a machine learning capability that sat inside of officer networks in the DoD, learned about expertise, how that expertise was changing in real time and applied that new learning to seed the network as Jeffrey was describing with the right experts at the right time to solve an issue. So it’s a very exciting domain, it’s one that we feel is just getting started and there’s lots of room to grow.
JEFFREY DAVITZ 03:43
Think about it as flipping the problem. There were some talks earlier about humans augmenting algorithms – so human input – people do this with clustering, seed the cluster in some kind of way. We’re actually flipping it on its head, what we’re doing is it’s machines which are actually augmenting people. What they’re doing is actually enabling social networks to solve problems better, because ultimately it’s people who solve these problems. We’re all accustomed to these – in some cases – incredible achievements of machines being able to answer questions, you have automated algorithms that can somehow extract out the pattern. But generally speaking for most of the problems that I bet most of you are concerned about it’s really ultimately people who solve those problems and networks. So the question is – given all this data that you have how can you make this network do that more effectively? So it’s the data and the machine learning which is augmenting the human problem solving.
JO MAITLAND 04:44
So more like Watson?
JEFFREY DAVITZ 04:45
It would be the flip of Watson.
JO MAITLAND 04:47
Or flip of Watson.
DAVID GUTELIUS 04:48
The really special part about this is getting the fine line between where the machine learning agents leave off and where the human network takes up. That’s a very difficult design challenge, there are a lot of efforts that we’ve heard about – some of them today – Watson is another example, where it’s taking in external signals of some sort and then pumping out binary answers or factual responses or things like this and that’s an exciting field. I think where we’re differing here in what we’re describing is this constant feedback loop that you design into this larger system where it’s taking incremental feedback from these humans in the network and incrementally changing and training and improving the way that the machine does what it does too. So it’s this virtual circle that we’re trying to design and insert into these networks.
JO MAITLAND 05:48
So how do humans need to work differently than in order to make that feedback loop successful?
DAVID GUTELIUS 05:57
That’s an excellent question and it goes back to the design challenge in one respect where what you want to do with any of this stuff is make that technology recede – recede from your focus. Because the focus you want to be problem solving on whatever humans are doing already, not on trying to make technology work for them. So part of the design challenge is creating a capability that does what it does very quietly in lots of ways, it’s unobtrusive and doesn’t get into the kinds of work flows that you really want to encourage inside of work groups.
JEFFREY DAVITZ 06:38
The fact is that people don’t adopt technology. By and large individuals don’t – this has been discussed – Gus was talking about this issue where Dave and I were back in the Intel world trying to talk to some of the people from the Intel world and they talked to some of the analysts and the analysts when they described some fancy new tool would do this. Which for someone who said well what is that and the answer was well it’s that, encrypted. Because they’re not interested in lots of new tools sitting on them, what they’re interested in is getting tools that are embedded in the way that they actually work.
JEFFREY DAVITZ 07:14
So what Dave was describing was an approach to design and one of the things since we understand a lot about the way that human social networks work – we do – I mean, we’re humans, we understand about how when you’re asked a question you have a few different choices about do I answer this or do I try to route it? That’s a very nat