E60: Celeste Baranski, VP of Engineering at Numenta – Interview

September 27, 2016


This is a great interview with Celeste Baranski. Celeste is the Vice President of Engineering at Numenta. Numenta is a small company with an interesting dual mission: to reverse engineer the neocortex, and then apply these neocortical principles to create intelligent machines. Numenta does basic neuroscience research and technology development, all of which is available in the opensource.

Celeste has a varied engineering background, from managing large teams to helping startups innovate as a designer.

We talk about what Numenta is working on, how Numenta is structured and Celeste’s role. Here are some other things we talk about:

-How does Numenta develop its technology?
-What kinds of things can Numenta’s technology do today?
-In 5 years what type of problems could Numenta be solving?


David Kruse: Hey everyone. Welcome to another episode of Flyover Labs and today we are lucky enough to have Celeste Baranski with us, and Celeste is the Vice President of Engineering at Numenta. And Numenta is a Machine intelligence company. So what does that mean? Well, their intelligence can help companies analyze IT and Server Data or anomalies and stocks or rogue behavior within a network. So it’s quite interesting. And Celeste I think is in charge from the engineering perspective, taking the algorithms and intelligent and building the architecture and framework to make it useful, but we’ll find out more about that. So I invited Celeste on the show because I’m curious how she makes this happen and just to hear more about hear background. So Celeste, thanks for coming on the show today.

Celeste Baranski: My pleasure.

David Kruse: So yeah, like I mentioned I want to hear more about background. So maybe – you have quite a rich background, so could you start telling us a little bit about what you’ve done in the past. That would be great?

Celeste Baranski: Sure. I got a Bachelor’s and Master’s degrees in Electrical Engineering from Stanford and I worked my entire career here in Silicon Valley; most of it in startup companies. I’ve been in seven startup companies and most of them working in very early leading edge consumer products; so most of them not successful, but a lot of fun, a lot of interesting technology. And then I moved, I did hardware design. I was a hardware engineer for the first few years and then moved back and forth between managing small groups and design and then maybe after about 10 years I moved fully into Management. So I’ve been managing software and hardware teams in Silicon Valley, on again mostly consumer products up unit Numenta.

David Kruse: Okay, and how did you get involved with working with startups and why did you continue down that path?

Celeste Baranski: Well, that’s a good question. So I joined a fairly large company right out of Stanford. It actually wasn’t that big. It was a few thousand people, it had just gone public. It was telecom and it was a little boring. I mean it was fun because there were a lot of new grads from Stanford who they hired and it was a bit like a continuation of college, but I didn’t feel especially challenged and so one of my – a guy I knew from my master’s program had gone to a startup called Grid Systems, which was working on a laptop. This was the early 1980s, so this was well, well, well before DOS even shipped, let alone the laptop was even conceived of. And it just, he showed me what they were doing and I was just fascinated, so I joined them and I just, I loved the ability to work on things where you really got an impact and a small company, a startup company, that’s what it’s all about and especially on things that maybe new and excited that not a lot of people are working on. So I just kept doing it.

David Kruse: Interesting. And so what did you, in that first company, that one was kind of working on the laptop, what was your role, what did you do at that company?

Celeste Baranski: I was a hardware engineering. So I went in and designed the whole, the board and designing the chips on the board and I managed the project with consultants. And for like a 22 year old out of college, it was an amazing experience.

David Kruse: I bet, I bet. I was going to ask you that, out of all those experiences, was there one that you thought was the most interesting or helped you the most or…?

Celeste Baranski: Probably a company called Handspring, which I don’t know if you’ve ever heard about…

David Kruse: I have heard of Handspring, yeah.

Celeste Baranski: Okay, I ran the engineering group there that was later in my career. So I joined in I think 1998 and it was just a fabulous experience. We did invent the first SmartPhone, and it was called the Treo. It was well before the iPhone was ever conceived of. And it was widely successful for a couple of years. Unfortunately the old dot com crash kind of caught up with it, caught up with us cash wise, but it was an amazing experience and just the product, the people, it was just really fun.

David Kruse: Interesting. Did you guys grow quite fast in those two years?

Celeste Baranski: We did. We went from – well at the time our revenue was the fastest growing ever for a consumer products company even beating out Sony. Yeah, we went from zero to $400 million year runway over night.

David Kruse: Oh my goodness!

Celeste Baranski: It was incredible; this is before the doc com crash. And that wasn’t on the cell phone. We had actually spun out – well the two founders of Palm Computing, Donna Dubinsky and Jeff Hawkins had left Palm and wanted to do this SmartPhone, but they ended up first doing another organizer. The Palm pilot wasn’t a very successful organizer. We did an organizer and it was, the organizer took off and really started our runway before we even started on the SmartPhone. So it was just a – it was a fun experience to design products and then I would see people carrying them, right and I would stand in line and three people in front of me would have our products on them. It was really fun.

David Kruse: So by then were you in charge of kind of overall hardware or what part of the…?

Celeste Baranski: No, at that point I was VP of engineering for hardwired and software.

David Kruse: VP of engineer, okay, got you. All right, interesting. And so let’s talk a little bit about what you are doing now which is interesting. So how did you get involved? Well it sounds like you were involved with another one of Hawkins startup, yeah, but how did you get involved with Numenta?

Celeste Baranski: So I actually met Jeff back at GRID which was the first laptop company I talked about and he joined about a month after I did at GRID. So we were both very early in our careers and he, from the very beginning we’ve been friends since then and he talked about his designer to figure out how the human brain worked, 30 years ago to me. And so I’ve always been fascinated with it, just incredibly fascinated. Then I worked pretty closely with him at Handspring and he really invented the SmartPhone as well as the mobile organizer. And he after Handspring, he went into Numenta. Well, he actually stated doing a pure research neuroscience institute first and then he fell in Numenta. About 12 years ago now, in 2005 and I’ve just been fascinated even since listening to him talk about it and then hearing it over and over again. So I was – I really wanted to do something about it, but I’m really a hardware engineer, I’m not a software engineered. I ended up starting a company with a couple of other colleagues back in 2007 that we were the first developer on top of Numenta’s platform. It turned out to be too early and we had to shut that down in 2010 and then I was lucky enough to get a job here about two years ago when the original VP of Engineering here at Subutai wanted to go into full time research. And Don and Jeff had both worked with me and so even in my background, I’m not a Neuroscientist, I’m not a computer competition neuroscientist or anything like that. They wanted somebody who could do a lot of different flexible things and could manage a team though a company that has a lot of different changing priorities and different missions and so I got to join them and I’m just having a great time. It’s really, really fun.

David Kruse: Wow, that’s quite a past that’s interesting. That’s cool, and could you – I tried to describe Numenta, but can you maybe give a brief description for the folks?

Celeste Baranski: Yeah, sure I tried – I practiced doing this in a few sentences, because it’s actually pretty complicated.

David Kruse: Okay.

Celeste Baranski: We actually have a dual mission and this is – I’ve worked in a lot of companies in Silicon Valley and I know a lot of them. This is the most unusual company I’ve ever worked at. We have a dual mission; one of them is a neuroscience mission, which is to reverse engineer the neocortex, which is just incredible and fascinating and that is probably our number one mission and so we really focused on it first. But then the second part of the mission is to apply the principals we’ve learned about the neocortex to machines or to computers, which is where the machine intelligent piece comes in. So we are working on neuroscience research and theory and algorithms and publishing those algorithms and then we are also publishing example applications of software that used the algorithms in real world situations.

David Kruse: Got you. So you do you have a number of researchers?

Celeste Baranski: We have – our whole technical team is 12 people. So the company is about 15, it’s pretty small. These were insider funded, so we don’t have outside funding just so we can concentrate on what we want to do, so we stay small. The 12 Subutai who is the VP of research here and one other PhD student or PhD graduate, they have the titles researched, so they do full time research along with Jeff Hawkins and everybody in engineering who works for me. Also I’d say everybody contributes. It’s really a gray area here where you do something for research. Everybody is a software engineer, except me. So there is a lot of back and forth and we focus on what’s important to move the research forward, really.

David Kruse: Interesting. Yeah, that’s unusual for like a smaller company that had research first focus. That must be fun.

Celeste Baranski: It is; its total fun.

David Kruse: Yeah, so how do you balance that between – I mean like you said research comes first, but you probably want to make money to at some point or another soon.

Celeste Baranski: Well. So we don’t have any revenue goals right now.

David Kruse: Okay.

Celeste Baranski: So we actually don’t concentrate on making money. What our mission is as far as the technology goes is to get as broad use as possible of it. So everything we do is in the open source, which under the AGPLV3, which is a license where anybody can use any of our code to do non-commercial applications. So researchers, any kind of research within companies using it to try something out, proof of concepts, you don’t need to worry about it. If companies do want to use it commercially, we do offer a commercial license which we have a few partners who have commercial licenses, but our focus in not necessary on raising a lot of cash right now. It’s getting the technology valued and to be of use in both academia and business. So we are really lucky to be able to do that, because you know the way our business is set up we are in venture capital backs, we don’t have investors waiting for the return and within Egger and Jeff and the few investors who put in some money over the years really are in a very long term focus in figuring how the brain works.

David Kruse: Interesting. I mean that’s probably where a lot of kind of breakthrough technologies will happen too, when you have that kind of devoted focus on deep problems, not just trying to spin out something that can make money in six months, which most companies are focused on. But they are going to miss out on the kind of the deeper game changing technologies.

Celeste Baranski: Yeah, that’s what I find so fascinating with this company is most places like this are like IBM or Xerox PARC and they have the money to be able to do these kind of research, but then they are big companies. So the bureaucracy is on top of it and we don’t have the bureaucracy, yet we have the ability to really focus on research and we have a small team, but is also, it’s incredibly productive. So it’s really a very interesting mix and I think doing really important work.

David Kruse: And can you explain the overall technology and kind of the idea behind how it works?

Celeste Baranski: Sure. So just second, let me see how I can do this pretty quickly. So we it sounds that it’s how we work. Some of your other questions were about how we work here, so it’s kind of mixed in that. So we do – we are working on a overall framework of a theory of how a biological intelligent, which can also be applied to machines. So if you think about what intelligent is, there is really only one intelligent system that everybody agrees is intelligent, which is the neocortex. It’s mostly humans, but you know all mammals have a neocortex, so that really pretty much everybody agrees that’s the seed of intelligent. So we go from that standpoint. We are trying to understand what the brain is doing as a system. So how the neurons interact to learn patterns, which is a part of in part of intelligence. How the sensory motor has come in and how they influence the pattern. So we do a lot of real basic neuroscience learning, which is reading papers, it’s collaborating with other academics and it’s coming up with a theory of how everything works together. That’s done mostly in small group meeting. Jeff leads those and the researchers lead those and they talk through things and sometimes it moves slower than other times and you know the methods been around for 12 years and we’ve made some progress but we are not done. We think we are making a lot of progress on the theory over the last say six months, which is great, but we aren’t done yet. Then once people are feeling pretty good about theory, we start – well, we test is out by looking for papers in neuroscience literature to either support or maybe not support what we are thinking. When we get closer we start to implement some algorithms and software to see if we can make the – apply the principals to data in a research framework. And once we get, if we get those working, then we actually write the algorithms and production code and release them into our platform. So the platform is called NUPIC, Numenta Platform for Intelligent Computing. It has algorithms that are in the open source that people can use for doing commercial activity, but it only represents maybe 40% of our theory right now, because it’s the stuff that’s more hardened and tested.

David Kruse: Got you, okay. And how does it – I know, I read – and this is long ago, so I read Jeff’s book on intelligent and you know it kind of talk about the structure, kind of a hierarchical structure. Can you kind of talk about that a little bit?

Celeste Baranski: Okay, about the brain architecture sense?

David Kruse: Yeah.

Celeste Baranski: Okay, so it’s a known neuroscience fact that the neocortex is uniformed both physically, so the neurons in there looked largely similar. The regions are largely similar, so if you sliced it opened it’s in any place, it’s pretty similar and neuroscientist have also known from many decades that even if you have certain areas in the neocortex that are attached to a sensory rate. So there is a vision area and auditory area, but if something happens to that area, other areas of the brain can take over those functions. So the brain is pretty adaptable and flexible, and so what we’ve been working on is trying to figure out how a group, a region of neurons in that area learn patterns and so our premise, the theory says that the synapses on the neuron are as they grow over time, that’s how learning is done. And so our software implements in not artificial anurans. So the artificial neuron nets and are very different from what we are doing. That may have been biologically inspired at the beginning, but then the artificial neurons are very simple and have nothing to do with biology. So what our model does is not exact. So it’s not Neuromorphic; it’s not trying to mimic every single piece of brain, which there are people trying do that, but that’s super complex. We are talking what we think are the principle, the important principles for learning and then implementing them in software. So the neuron, our neuron model has a number of active dendrites that are modeled that come in from different areas and they take date in the form of a sparse distributive representation, which is we think is the language of the brain, it’s what we are use to, which you can think of as a large sparse sector using a brain. There are billions of neurons, but only a very, very small fraction or ever active at one time and they are kind of distributed all over your brain. So the data that we take in mimics that function. And then those, the neurons that we implement in the region, take those SDRS in and they can learn patterns over time mimicking the synapse growth. And then once you learn a pattern and you learn it without having to have lot of training data and label data, its learning all the time just like your brain.

David Kruse: And yeah, that’s what I was curious about. I mean you might have just answered it. But I was curious how it’s different in neuro networks. I mean that’s one thing that often neuro networks need a huge amount of label data. But how is yours – I mean, do you have different algorithms or how – obviously you do, but how is it set differently?

Celeste Baranski: So, to think that we – we have a nice blog post on this too on our website. But we – the way – we get asked this a lot by everybody that comes in. So we’ve kind of defined it. An artificial intelligent is in an incredible period right now. There is a lot of exciting stuff going on in all kinds of areas. But it’s also kind of gray because it isn’t – there is not a lot of. It isn’t really a mature field yet, so even the terminology is really hard to understand. So we kind of defined things into three categories. One is, types of artificial intelligence. One of them is traditional AI, which you can think of. It’s been around for many decades as more of an engineered solution. So engineers saw you know this is the way I want – I want something to act like this from an engineer, an exact solution or a rules based system or you know rules based system is probably the best way o describe it. You know IBM watching is actually a good example of traditional AI where they’ve engineered this amazing computer to do very specific tasks. But it’s not really flexible and it’s not intelligent in the sense it doesn’t learn over time; so that’s one class. The next class you can think of as mathematical or artificial neuro nets and those, it really is a mathematical or/and statistical based sciences, which is and it works great. There has been a lot of breakthroughs. Even those artificial neuro nets are around for I think 20 years, maybe longer, the amount of data available and the computation and power that are available over the last few years, there have been a huge amount of breakthroughs in this. So deep learning is in this category and it does really well at classification problems. So if you can provide lots and lots of data, you can train these highly complex mathematical models to do a really good job at recognizing faces or recognizing specific patterns on the web and it’s really valuable. I mean a lot of companies are doing research into this. But it’s not a biologically based approach and it doesn’t learn continuously and it doesn’t really take into account patterns that change over time. And so I’m getting under the biologically, biologically constrained algorithms, in each strand there is one of them. There are others that do this, but we are probable the only commercial one and the things that distinguish us against other AI’s is we are all based on temporal patterns. So patterns would change over time, because you know as your brain reacts to things that happen over time and isn’t a static batch process. It also learns continuous. So the learning never stops, there is no training period. There is a period when you are learning it like a baby learns something, but then you learn for the rest. Learning is never turned off and you continually learn till patterns change, you learn the new pattern. So you don’t go back from training to operation. Those are probably the two biggest changes and we think this is really the true, and if you define intelligence as being able to learn it, that the biologically possible algorithms are really the ones that are intelligent. That’s not to say that deep learning doesn’t have any venue, it certain does, but it isn’t – since it doesn’t take it into effect, into accountable data and by the way there is like 30 different deep learning algorithms that engineers have to piece together for any one particular problem, where the biological ones we have two or three algorithms and they are used for everything. So the code base stays the same, the data.

David Kruse: Wow!

Celeste Baranski: Yeah, it’s petty amazing. Now you do need – depending on what kind of date you are feeding in, you need to encode the data into a sparse dispersion representation. So there is a process, the encoder process that does change depending on data type, but the algorithms itself looking for anomalies and today they look for anomalies and can do predictions, don’t change at all. So it’s a definitely different approach.

David Kruse: Interesting. It’s almost like you have adaptive learning kind of built in.

Celeste Baranski: Its – we call either continues learning or online learning, but yeah.

David Kruse: Interesting, wow! Okay. And so for your role what – I’m curious what you kind of do, what your priorities are as VP of Engineering and then yeah, we’ll start there?

Celeste Baranski: Okay. So this is, it’s a really different job, because we are pretty small in money. I’ve run groups up to, gosh, I think at Palm, after that we got bought by Palm and Handspring and my group is almost 500 people, very, very different role here and we don’t – since we don’t do commercial products. So I do a mix of project management. If we are doing some projects, I do hiring. We actually don’t have any full time positions, but we have a very strong intern program. So I help, I manage the intern program with my Director. I do writing, I help with the business stuff, I work with the marketing person. I pretty much do everything I can to keep the group motivated and working on the highest priorities and focused. And I try to get Jeff and Subutai they are trying to be able – they are concentrating on the research, because that’s the important stuff. So I do whatever I can to motivate the team to help them.

David Kruse: Got you and when – okay so you have at least two or three algorithms. And then how do you from an engineering perspective make it easy for anybody to kind of tap into this intelligent?

Celeste Baranski: So we have all our code in source. So this new pick is an open source project and our production algorithms are in there. We also open source all of our research code and our research activities also, so people can look at what we are doing. We try to be super transparent. We do have – so we do have patents, we have over 34 issued patents in this area now. We have a non-assertive – we’ve made a non-assertive statement that we won’t assert our patents against people doing academic research or using them for research purposes, but we do, we feel our value as a company is largely tied up in those patent. But after we – but once we apply for patent we put everything into the open source so people can see what’s going on. And we have an open source community, probably about I think 400, 500 people now worldwide who are doing various things, trying to use the code. We try to keep it documented. We don’t have a formal SDK or super formal documentation where the code is fairly well documented is a good wiki, but we haven’t had – we don’t have the bandwidth to do that. So the community does some documentation. We have an open source coordinator who works for me and Matt Taylor and he does, he has been focused lately on education. So he is doing a YouTube video series on called HTM School. HTM, Hierarchical Temporal Memory is the name of our theory. So we try to support the community however we can and then try to get them to support each other.

David Kruse: And can you give an example of, well ideally more cutting edge the better, but some of its probably confidential. But an example of, you know I listened to some used cases at the beginning in and around stocks or monitoring rogue behavior. You can use one of those or if there is another one where you know your algorithms learned and helped whatever the used case might be.

Celeste Baranski: Okay, I can tell you about a commercial one today and then maybe kind of far reaching one when we get further. There is in one of its business models, 50 years Numenta did develop a commercial application that was called Grok that used the learning algorithms to detect anomalies in server farms. So we ended up writing it for AWS, so you can monitor your servers in AWS and there is something, you know when an anomalies happens you get notified and we’ve, I think we have proven pretty well that we can find anomalies much early that many other techniques like threshholding. So we, about a year ago we had a deal with another company and it was a startup, I think they are going by the name Grokstream now and we transferred that whole application to them and they are selling it commercially. So that is the part of the biggest commercial application that’s using our technology today. So its finding anomalies and service groups.

David Kruse: And what will be the example of an anomaly that you would find that would be a problem or…?

Celeste Baranski: So for instance you will have. So in an AWS server you probably get 500 metrics, every five minutes you get measurement out of 500 metrics that will tell you how your servers, how your server farm is doing. There is the CPU usage, there is network traffic, there is low balancing, so there are all these signals. So if somebody – for instance we run this in our, in the monitors in our development process. So the last couple nights ago, over the weekend we got an anomaly on one of our servers and it turns out that one of the engineers kicked off a manual build process versus having the normal build process that’s in the pipeline. So if you are in a – that doesn’t matter for us because that’s okay, but it you are in a – trying to monitor large IT infrastructure, people can’t keep their eyes on so many things at once. So you need a filter to tell you when something unusual happens and it was a combination of CPU metric and network in I think, and it just gives you a red flag and says, hey, something unusual happened with this server. Then the IT professionals can go there and see maybe it’s something bad, maybe the server is starting to run out of memory or somebody is doing something, somebody is accessing a file they shouldn’t somewhere.

David Kruse: Interesting, and that’s because algorithm was trained essentially, what’s kind of a normal day in the use case.

Celeste Baranski: A normal pattern.

David Kruse: And then when the pattern is disrupted or yeah, there is an anomaly then it triggers something, okay yeah.

Celeste Baranski: And its trained on a temporal pattern. So pattern over time, so it’s more than just a threshold or something that was, it was told to look for, things that learned itself over time.

David Kruse: And so it doesn’t need like a lots of labeling necessarily, like would it use kind of like…

Celeste Baranski: No, not necessarily.

David Kruse: Really? Interesting.

Celeste Baranski: It just learns what it sees. So you can add labels on it in an outer loop if you want to label some of that as good. What this pattern really does is it looks for changes and that’s what our – we have a little example application I think you might refer to as HTM for stocks, which you can download on your phone to try to see what HTM looks like in a consumer application and it monitors stock volume, stock price and the number of tweets, the Twitter volume for 200 large cap companies that have traded on the stock exchanges in the US. And it tells you when there is an anomaly for any company and then it sorts, on your phone it sorts the list, you can see where the anomalies are. Now it doesn’t tell you whether that anomaly was good or bad, it just is a change. So if you wanted to, you could add. We’d love for somebody to take that App and add things around there and they can filter out and classify whether the anomaly is good or bad. But it’s just, it’s a change in something.

David Kruse: How much more, with that example, could you add 20 more data sources and would you be able to handle that or…

Celeste Baranski: Oh! Yeah. The reason it has 200 is because that is – since we are running this for free, it turns out you can get, basically you can search 200 Twitter streams for free in a pubic API from Twitter, otherwise you got to pay a lot of money. So we are basically just using this as an example of what you can do.

David Kruse: Got you, and I almost felt like new data sources like you know right now you are dealing stock prices and Tweets, but what about if you wanted to bring in like lots of other…

Celeste Baranski: We thought about putting in new sources that you could, yes.

David Kruse: So you can bring.

Celeste Baranski: You could actually do that, yeah.

David Kruse: Interesting okay. Well, so it’s quite flexible. Do you have another example maybe down the road?

Celeste Baranski: Yeah, sure. And one other interesting one that one of our other partners today is doing is, when you think about its pretty huge, its natural language processing. So when you think about, there is a lot of work to be done in language processing, but it’s really a learning process of what and understanding what words actually mean, not just a big matching thing that can be done. So I think Watson does a good job of doing natural language queries, because they just have a massive ton of information, but I question whether Watson really understands the sentiment of something’s been said. I don’t think that sort of thing. But if you can really understand what something means, then it opens up a whole new way of dealing, computers dealing with the language. So Cortical.io is using HTM to do that, to really understand natural language. So its again the same thing. You can apply the same algorithms to a totally different data source that’s not even a number.

David Kruse: Interesting. Wow! Yeah, yeah.

Celeste Baranski: And then if we look in the future. So what we are working on now is in the theory side is trying to understand how the sensory motor behavior, the sensory information comes in and how that interacts with the neocortex and the neocortex then directs behavior, so that your body is really directed by your brain to do things, but your body is what’s seeing and touching and hearing things and taking the data to the brain. So it’s all big circle and that is kind of the complete theory of the neocortex once you get that circle done. Jeff has a real neat example. He has been talking about – unless he talks about you can – you know one thing that could be a learning machine that could send into outer space, because if you think about, you don’t want to send a bunch of people to Mars even though we saw the Martian movie whatever. You know it’s when you first go to a hospitable planet people can’t be there, because you know its kind who is the first people that goes there and how do they even live. So if you could send a machine there who had learnt how to do things like building structures, but was intelligent enough to adapt when things happen, if there is a windstorm or they run out of gas or whatever they are going to – they are going to be able to do things that humans can’t do even. And I thought that was a fascinating example, because it can change the way mankind lives in this world and even in the universe.

David Kruse: That’s true intelligence if we could go that way.

Celeste Baranski: Yeah, and its good and it’s not an evil thing either right. All technology can be used for good and evil and certainly intelligent machines, you can imagine somebody programming them or making them for evil, but there is also incredible good that can come out of them.

David Kruse: So in five years what would it be, this is a kind of tough question. But in five years what would be an application or something like hey, if we could do this by this time, that would be pretty amazing.

Celeste Baranski: So, that was a hard question. But I’d say if you could go back when I talked to Jeff when he was first starting Numenta, he didn’t think it was going to be this long. You know if he had looked in the further like 12 years later and I’ll be doing this, I don’t think he knew that, so we don’t know. Certainly if we can, if continue the progress that we’ve been doing over the last nine months in the theory, we could be doing things that do integrate sensory motor behavior. The most obvious example is really good robotics, but there is other things like what I mentioned about the space basis version. So five years is one those strange timeframes where I can kind of see where we are going be in a year and what we might do, but five years, I don’t really know. I know it’s going to be exciting.

David Kruse: Yeah! Well I just like to hear exciting stuff, even if it takes you fifteen years, that’s okay. Maybe I should have said 15 year and then it’s so far off everyone would have forgotten by then.

Celeste Baranski: Yeah, I don’t know yeah. I think if certainly we hope in our lifetimes that – I really feel like one of the reasons I’m working here is I do think that Jeff, what he is doing is brilliant and I really think it’s important and us and whether HTM and this implementing we are doing is exactly right, I don’t know, but the concepts and ability to understand what the neocortex is doing will impact mankind. I just think that’s fascinating and I hope it’s in my lifetime.

David Kruse: Yes, yeah, me too. I’m so glad you guys are doing this. Because like you said, I mean there is so much research around neuro networks and I think people are trying to advance that, but it doesn’t learn – it’s a lot of batch processing and then labeling and so you guys are, like you talked about doing more of a advance work that needs to be done in order to go the distance.

Celeste Baranski: Really, it’s the understanding of the really, the system the brain works and there is a lot of neuroscience research, but it tends to be very narrow and deep. You know the mapping, the neuro map of some small animal or…so that’s all great and its really interesting, but really understanding how the system works as a biology and then applying to other things is just amazing.

David Kruse: Well, unfortunately we are almost out of time here. But so another question I had around Numenta was, I mean it sounds like it’s very good for identifying anomalies. So any place or any situation use case where there could be a potential for anomalies, where there’s lots of data. It sounds like your HTM framework could be quite appropriate, is that fair?

Celeste Baranski: Yeah, I think it is and its over time. So streaming data is incredibly important. So a big batch file is not going to bring in value to that. So the internet of things, a really interesting application area. Lots and lots of sensors, streaming lots and lots of data that you can’t train models for; that’s a really interesting area.

David Kruse: Is anybody using it in that space?

Celeste Baranski: We don’t know, we have lots and lots of interest. I guess we are open source. There could be a lot of people doing things that we don’t know. We don’t have a partner that is formally that we’ve engaged with is doing that.

David Kruse: No, but I mean that’s a big thing with IOT is that like you have all these sensors and then what are you going to do with it all and so you are going to have to build like custom models for every single factory and every single company. But if you could come in at least and provide some initial intelligent for one of the sales cycle, it would be a lot easier, but then also yeah, you’d be so much further ahead. That’s interesting. Okay, well I think that’s just about does it unfortunately.

Celeste Baranski: No, that’s okay. It was fun to talk about.

David Kruse: Yeah, I know, I really appreciate your time Celeste. I didn’t know your – I knew your background was interesting, but I didn’t know it was that interesting. So I was glad to hear what you got and…

Celeste Baranski: You’re using the word interesting and not strange.

David Kruse: Yeah, well they both kind of go together sometimes. No, I really appreciate it and I’ll be curious to see how things go and I think at least I personally wanted to check out the platform a little bit more and yeah, so thanks Celeste for coming on the show.

Celeste Baranski: Your welcome. Thanks for having me.

David Kruse: And thanks everyone for listening to another episode of Flyover Labs. Definitely appreciate it and I hope you enjoyed it as much as I did. We’ll see you next time. Bye everyone.