Editor’s note: WRAL TechWire is teaming with YourLocalStudio.com, a 10-year-old self-described video agency, to publish what we believe is a riveting, deep, deep dive into the world of Artificial Intelligence. Alexander Ferguson, the CEO and founder of YourLocalStudio, and his team have done a series of in-depth video interviews with thought leaders in the AI field – many of whom are based in the Triangle. Here is the latest in the weekly series, an interview with Robbie Allen, CEO of Infinia ML.

Welcome to the UpTech Report series on Artificial Intelligence. I’m Alexander Ferguson. This video [and interview] is part of our deep dive interviews where we share the wealth of knowledge given by one of our panel of experts.


For our second series of deep dive interviews, I sat down with Robbie Allen, CEO of Infinia ML in Durham. Robbie has a degree in engineering and management from MIT and is completing his Ph.D in computer science at UNC Chapel Hill. He owns six patents and has authored eight books. So he has a great deal of knowledge in this space.

  • The interview: I asked him about how the recent boom of machine learning has affected AI overall, how he got interested in this field and how these technologies are being applied in business today.

What’s happened is that machine learning has been included as a part of some of the other aspects of artificial intelligence like computer vision, like natural language processing. And so we’re seeing sort of a renaissance in many of those fields of artificial intelligence. In large part thanks to machine learning as an underlying technology driving them.

  • What’s the difference between machine learning, deep learning and reinforcement learning?

So machine learning, at sort of a fundamental definition, is the way I describe it, is automating something and learning patterns with data. So essentially you start off with the data set and it’s essentially learning, the software’s learning patterns in that data so that when you give it new data in the future, it can make a prediction or it can tell you how it’s similar to what it’s seen in the past. So that’s really what machine learning is.

Deep learning is a specific type of machine learning that you can think of as just a more complicated version. It allows you to really go at a deep level, to use the term. To find patterns in a more intricate way than kind of, maybe what we’ve done in the past. That’s primarily due to the computational resources that are now available. As well as the enhanced sets of data that are available.

Now we can apply essentially more processing, more capability at traditional machine learning algorithms so that they can find patterns in more nuance ways. So reinforcement learning is a way to train an algorithm based on a reward system. So essentially, can you identify a behavior that you think is good and what happens is the algorithm will trial lots of different approaches and once it gets rewarded, it then files that away.

Okay, the path I took to do that was the right path so I’m gonna keep track of that and I’m gonna try to do more of that versus something that was not a success. You know, with natural language that’s a special case that, really there’s two main fields within machine learning, natural language and then vision, which is image processing. Those are the two main data inputs that you get.

There’s also tabular, numeric data that you could also work on but often times you’ll hear about natural language processing or computer vision as the two main areas. Vision has probably had the biggest advancements to date. And you’ve maybe heard the most about it, whether it’s facial recognition or things of that nature. Just because it’s a more constrained problem space, right. You’re talking about an image and really you’re kind of trying to focus on identifying, maybe things in the image.

Versus natural language, tends to be a little more unbounded. And again, there’s multiple languages, there’s all sorts of nuances. Two language that are not present with images. And I think it tends to be a harder place to kind of get the same level of accuracy. And then when you start talking about generating text, that’s a whole ‘nother game.

My first company, Automated Insights, that’s what they do. Is automatically generate texts and it’s decidedly not machine learning based. At least the text generation piece, because of this probabilistic, sort of attribute of machine learning. We can’t deliver results that are right 80% of the time. Because people are so finely tuned to any issues in text they can easily identify when there’s something outta whack.

An image actually can have a little bit of error built in and people be okay with it but if your sentence comes out garbled one out of ten times, you’re gonna know it, and you’re not gonna appreciate that. And so, natural language is a little bit harder case in the machine learning space just because people are very finely tuned to text. If something’s off, you can spot it very instantly and it doesn’t sound very good.

  • What is natural language processing and how does it work?

Traditional natural language processing techniques are what are referred to more as brute force techniques. That is, you try to definitively describe all the different parts of speech for example. And you’ll have these very complicated rules based systems, the heuristics that are kinda built in to try to automatically determine all the different parts of speech. But the problem with speech and language is that it’s changing all the time.

Just what is machine learning? AI experts talk definitions and uses in UpTech series

It’s a very tricky thing to really define discreetly. And so, with machine learning, when it came to prominence what that allowed you to do is actually automatically learn the rules of language versus someone actually having to define all those by hand. Now you can just throw a lot of data at it and have it try to automatically detect the patterns and detect the rules of speech and the rules of language.

Traditionally you may try to build a natural language processing system off a traditional, voice based capabilities but guess what? Everybody says things a little bit differently. I have my own accent. I’m from the southern United States and you can probably hear that a little bit. So traditional natural language processing and speech processing techniques would fall apart because they would have to be built off of a particular type of speech system.

They weren’t able to really handle the nuance and the accents that were present. Now with machine learning based systems, it can be built with initial base model but it can also learn over time. In fact, Alexa and some of these other systems have capabilities where you can actually talk to it and have it learn your specific ways of speaking. And that just wasn’t possible before.

  • How did you get interested in AI and specifically machine learning?

It really starts back all the way to the mid 90’s when I started working in computer science. That kind of continued on to my first job which was at IBM’s networking hardware division and then followed by working at Cisco. And so I’ve always, kind of in that my early part of my career involved in this sort of networking and hardware space.

But even back then, the core of what I did was automate things. And really the way you can think about what is artificial intelligence really hope to promise is just an enhanced ability to automate. So I’ve been fascinated with abilities to automate things from many years back and now with artificial intelligence. And I did some of my work at MIT looking at artificial intelligence in the mid 2000’s, even then it just still wasn’t at a point where you could do it at scale and have the kind of impact that we’re seeing today.

Artificial Intelligence and you: Introduction to a new video series ‘UpTech’

It’s only early 2011 and ’12, again there was a confluence of things that happened that now makes machine learning a very practical tool to use when you’re trying to solve certain business problems and that’s just, like I’m a kid in a candy store. Because now it opens up all these different things to automate that just weren’t possible before.

  • How do you go about solving these business problems with machine learning?

The initial part is really translating the business problem to a machine learning problem. Because it’s not always the case that somebody tells you, alright we have this invoice processing system and it’s highly manual and repetitive and now we want to automate it. Okay that’s good and that’s a good frame to start with but you then have to translate that.

Deep dive: Artificial Intelligence can take over ‘things humans should arguably not be doing’

Well, what does that entail in terms of a machine learning problem, like what are the ways that we’re gonna apply machine learning to maybe break down that business problem into a series of machine learning problems that we can then ultimately stitch together to deliver a solution. So I’d say the first step is really helping define the business problem in machine learning terms.

Coming up

Stay tuned as we share the full deep dive interviews we had with each one of our panel of experts and our upcoming episodes focused on specific topics that will transform the way you think about artificial intelligence. All this on UpTech reports new series on AI.

What are applications for Artificial Intelligence at your business? UpTech takes in-depth look

Inside Artificial Intelligence’s disruptive power: From marketing to resurrection of characters, singularity, more