Spatial temporal analysis research projects are underway at IBM’s TJ Watson Research Center. Mudhakar Srivatsa, Distinguished Research Scientist, and Raghu Ganti, Research Staff Member, provide an overview.
Advanced algorithms and streaming data track and predict the behavior of moving objects, from asteroids to rhinos.
Adrian Bowles: We’re going to be talking about AI today, with a couple of researchers from the T.J. Watson Research Center, at IBM Yorktown. And, if I can have you introduce yourselves first, and then just talk a little bit about what you’re doing, then we can get into the details.
Mudhakar Srivatsa: I’m Mudhakar Srivatsa. I’m a Distinguished Research Scientist with the IBM T.J. Watson Research Center in Yorktown Heights. I have a small team of about 10 to 12 people working on distributed AI. The primary focus area under that is spatiotemporal analysis. Basically, we analyze patterns of moving objects ranging from people, to cars, to asteroids, to rhinos—something that we’ll talk about in a lot more detail in the next few minutes.
Raghu Ganti: My name is Raghu Ganti, and I’m part of Mudhakar’s team. Around about eight years back, we started the work on spatiotemporal analytics, and which has now grown much larger. We have several products, about 12 different IBM products, that we’ve integrated our library with. We are continuously working on more and more real-time problems as well as AI-related problems, on top of spatiotemporal data.
AB: You mentioned asteroids, and saving the world. Talk about the scope. You’re dealing with almost anything with geospatial data, or asteroids to rhinos is a big, big leap.
MS: Yes, that’s a very interesting question in itself. Automatically, it’s the underlying mathematical model that we are working on, so we work on an ellipsoidal mathematical model, so we treat the (that’s an ellipsoid), and we track movement of objects, whether it’s cars or rhinos, on this ellipsoid. It just so happens that asteroids do also take elliptic orbits around the sun, so the same kind of math applies there, except that in one case we are working on a geocentric coordinate system. We are working on latitude, longitudes. In the case of asteroids, we work on heliocentric, meaning sun-centric coordinate systems, meaning sun is zero, zero, zero in some sense.
RG: It’s about showcasing that our library is capable of handling on just very, very fast moving objects bound by physics laws, but also objects that have a mind of their own, and they’re moving around.
AB: Okay, so there’s a commonality in the modeling.
AB: And there’s a lot of machine learning involved. Can you talk just a little bit about that, and we’ll get into the details in one of the examples.
MS: In the case of say, protecting the rhino from poachers, it’s a very interesting machine learning problem.
Normally, you would approach by tagging the animals. You want to track their animal movement. We can track vibrations. That will tell them whether they are moving fast, moving slow. We can have GPS sensors on them that can give us the position of information, and so on. But the issue that when we put a tag directly on a rhino, is that the radio signals may be triangulated by a poacher. So, if you put a tag on a rhino, you might in fact be telling where the rhino is located to somebody whom we are trying to protect the rhino from in the first place. We have to put tags on other non-endangered species, which happen to be in the vicinity of rhino. They are sharing the ecosystem with the rhino.
In this particular case, we are putting the tags on the impalas, the zebras, the wildebeests. They are denser, they are larger in population, they are in the vicinity of the rhinos, they do tend to behaviorally respond when there is a poacher in the vicinity.
We do, in data sensing, meaning we don’t directly tag the rhino, but we tag these other animals and then we try and infer whether there is a poaching activity that’s going to take place in the next few
minutes, five minutes, ten minutes, fifteen minute kind of a time range.
AB: Because it has to be predictive to be useful.
AB: When you said infer, are you talking about inference in terms of logical reasoning, or in terms of when you take supervised learning, and you do the training, and then you infer from that. How are you drawing these inferences?
RG: It is about being able to marry some of this very little label data with more unsupervised techniques.
Things like clustering, and being able to identify what is the right label, and how do I train a model with very little data. That was the main challenge that we faced, and we all came that we’re using
unsupervised techniques with guidance from the label data. It’s just saying that, “Oh, when a poacher attacks, the animals behave slightly differently from when a predator attacks them.”
RG: And being able to use that information and create AI models based on that, I think that’s the key in coming up with the right AI model in this scenario.
AB: I love the idea that you are modeling this ecosystem, if you will, by directly observing other participants, rather than endangering the rhinos as you do it. This is great, because it sounds like you’re combining … integrating multiple types of machine learning with sensors, what we think of as the IoT, to create a solution to preserve the wildlife. We have rhinos, and we have asteroids, are maybe the extreme. No disrespect to rhinos, but in terms of speed, when you’re talking about geospatial, it sounds like one of the issues is how fast can you react, but I would think that that is influenced heavily by how fast the object is moving.
AB: Can you give me any other examples?
RG: Another recent example is about being able to predict the encounters of flying airplanes, especially with the increased volume of the number of airplanes in air spaces.
A key challenge, and a very difficult problem is for the air traffic control to tell the planes what is the landing patterns, what is the takeoff patterns, and so on. And to assist in air traffic control, we want to be able to provide a predictive model that can tell you are these two aircraft, are two aircraft going to be where they should not be in the same joined airspace.
That is where I think last year, we developed this new operator for Streams that can do in real time, be able to ingest the volume of data that is coming in, the speed and velocity of which is quite large, because you’re monitoring this aircraft, and it’s in millisecond time ranges, to be able to identify and predict if these two aircraft maintain their current path, they are most likely going to end up in a situation which they should not be in, and alert the air traffic controller to be able to take evasive actions, much before they are able to even detect such a thing. That is another aspect of real-time data, and very fast moving objects, which is not as fast as asteroids, but not as slow as rhinos either.
AB: Sounds like a really practical application.
AB: I’m all for saving the rhinos, but I’m also all for saving myself when I’m in the air.
RG: Yes. Very much so.
AB: Excellent. Excellent. Well, thanks so much for joining us.
RG: Sure. Thank you.