Interview: Jennifer Marsman, Principal Software Development Engineer at Microsoft

Print Friendly, PDF & Email

In this podcast interview, I caught up with Jennifer Marsman is a Principal Software Development Engineer in Microsoft’s Developer Experience group, to find out about her experience at Microsoft and get her take on the upward trajectory of AI and deep learning that we’re seeing in the industry today. Jennifer educates developers on Microsoft’s new technologies with a focus on data science, machine learning, and artificial intelligence. In this role, she is a frequent speaker at software development conferences around the world. In 2016, Jennifer was recognized as one of the “top 100 most influential individuals in artificial intelligence and machine learning” by Onalytica. She has been featured in Bloomberg for her work using EEG and machine learning to perform lie detection. In 2009, Jennifer was chosen as “Techie whose innovation will have the biggest impact” by X-OLOGY for her work with GiveCamps, a weekend-long event where developers code for charity. Jennifer holds a Bachelor’s Degree in Computer Engineering and Master’s Degree in Computer Science and Engineering from the University of Michigan in Ann Arbor. Her graduate work specialized in artificial intelligence and computational theory. Jennifer blogs at and tweets at

Daniel – Managing Editor, insideBIGDATA


insideBIGDATA: Welcome to today’s insideBIGDATA podcast. I am Daniel Gutierrez, insideBIGDATA’s managing editor and resident data scientist. Today I’m here with Jennifer Marsman, a principal software development engineer in Microsoft’s Developer and Platform Evangelism group. Jennifer, it’s great to have you with us.

Jennifer Marsman: Thanks for having me, Daniel.

insideBIGDATA: Okay. Great. Well, let’s just start right in. So Jennifer, please tell us a little bit about your background and your experience so far at Microsoft. I see that you’ve come into data science on a rather direct route with your academic background in computer science. What’s interesting is I am seeing more and more people come into data science from other fields these days. So what can you tell us about yourself?

Jennifer Marsman: Well, I’m glad more people are coming to data science. No matter what your background is, it’s an amazing field which is just exploding right now. So everyone is welcome. Myself personally, I started in computer engineering, and then I did a master’s degree directly after my undergrad focusing on artificial intelligence and computational theories. This was 15 years ago, so a lot has changed since then, but it’s a very, very exciting field. I started with Microsoft right out of college as a developer doing natural language processing on a team that was essentially trying to build Cortana or Siri or Alexa 15 years ago. We didn’t have the cloud back then, so it was a little harder. I transitioned about halfway through my career, and I now work in the evangelism organization where I have the amazing job of going up to people and talking to them and just building amazing code samples and then kind of sharing all the amazing stuff that Microsoft is doing in the developer space. My passion has always been around machine learning and data science, so that’s where I tend to focus my work. I work both with customers as well as do some of my own projects in that space.

insideBIGDATA: Fantastic. It sounds like you have a great background for being an evangelist for data science. But can you tell us a little bit more about some of the work you’re doing at Microsoft?

Jennifer Marsman: Certainly. I’ve mentioned how I have some of my own projects as well as doing some work with companies to help them implement our machine learning technologies. So examples of some of the stuff I’ve done as just kind of for fun – one of the things I’ve been working on is using our cognitive services, which are a series of  pre-built models that you can call that are very easy to use because you are essentially tapping into a pre-built model that’s based on the decades of research that we have in Microsoft research. You just call REST endpoints, and then maybe send an image, and then get back a ton of information about that image. So for example, our computer vision API gives you back information on the tags, simple one-word descriptions of the image, as well as the caption, which is a little phrase about the image. It will tell you if it’s a black and white image, if it’s clipart, if it contains spaces, kind of what the dominant colors are, all of this wealth of information. Then the same thing with our face API. Again, I just send in an image and I get back the 27 facial landmarks that are part of the face, as well as the bounding box of the face, exactly where the faces are located within an image, as well as an estimated age of each face, an estimated gender, a whole bunch of information around facial hair and glasses and all kinds of other things. The glasses thing always cracks me up in particular. We actually had people wearing goggles in our training set, so the glasses thing is an enumeration between no glasses, reading glasses, sunglasses, and swim goggles. So if you have a use case where you need to know if someone is wearing swim goggles, that might have worked for you.

I was playing with the facial verification API, for example, which is our API which will allow you to train a model on images of the face, and then I can call it to see if any new faces match that face. I trained a model to recognize me. And then I’ve been working with a drone and have the drone take pictures of me or take pictures at a conference. I’ll have the drone fly, take images and then be able to recognize me inside of the images. So that’s just one of the kind of fun scenarios that I’m doing.

But I also work with a lot of corporations as well as inside of Microsoft to help people implement our machine learning technologies. One kind of fun story is around Xbox because who doesn’t love Xbox, right? The Xbox team is just – they’re so passionately devoted to their customers. They have a Twitter alias. It’s @xboxsupport. When something goes wrong and such, people can tweet at that alias and the Xbox team responds. A funny thing, they actually won the Guinness World Record for the most responsive brand on Twitter. I didn’t even know that was a thing but apparently, that’s a thing. So with all this data that they’re collecting, they wanted to create programmatic way to be able to mine it and be super responsive, because any feedback is a gift. So if they’re hearing things, they want to be able to take action quickly on what customers are saying. So they’re using our cognitive services, some of the text analytics capabilities in there, to be able to do some pretty neat stuff. So there’s a couple different parts of text analytics.

One thing it can do is language detection. Imagine taking all this stream of Twitter data that’s coming to this alias, this Xbox support alias, and sending it through language detection and being able to grab things in a different language. So let’s say I want to parse out English, for example, or Spanish, and see what issues that are happening by language. Then once I have the language, I can send it to various other services. For example, we have a key phrase extraction service in text analytics as well, and with that they can pull out the main point. So if someone says something like, “My service is down,” then it would pull out, “service” is what that phrase is about. So it can pull out the main points for you and make those really easy to find. We also have a service around sentiment analysis that gives you back a number between zero and one for how negative – which is the zero end – versus positive that that sentiment is of that text.

That is a fascinating thing because with all of these things together, they can take that data, put it through the key phrase extraction, and be able to pull out the most relevant things that people are talking about. So here are the key things, the key topics that people are talking about, and then be able to see, “Okay. Are these things largely positive or largely negative that people are saying?” They actually built this gorgeous bubble visualization with it, where they run it through those three services as part of our cognitive services. They get key topic extraction, and then they also get sentiment analysis. Then there are all these bubbles of, “Okay. These are the things that people are talking about.” Each bubble is labeled with one of those key phrases, and then the color of the bubble represents if it’s positive or negative sentiment. So green means people are happy about this particular topic, and red means that they’re unhappy. Then we also put the size of the bubble relative to how many people are talking about that particular topic. So a topic that a lot of people are talking about is going to get a bigger bubble versus a smaller one. So it’s a really neat visualization to kind of see what is going on in the Xbox world.

Another fun story – I’ve been working with them, and one of the developers got this set up. This happened right before the end of last year, and then on Christmas day last year, 2016, we  actually saw in the morning, a red bubble appeared and it had the word “code” in it. So that means people were saying something negative about a code. We saw that and right away we were able to click through and see the corresponding tweets that went with that. We realized really quickly that a subset of users were having trouble redeeming this special offering code that we had over the holidays. So by using that machine learning technology and the power of cognitive services, we were able to find that super fast and then resolve the issue really quickly to limit the impact we have. So that’s the story of how Xbox saves Christmas!

insideBIGDATA: Whoa, you are working on some really cool stuff. And I can sense by the way you just described it, you’re pretty excited about it. You got me excited. So I noticed that on your blog, there’s a picture of you wearing an EEG sensor on your head. Did you really create a machine learning algorithm to work as a lie detector? Can you tell us about this use case?

Jennifer Marsman: I sure can. So this was another one of those projects that I was doing for fun just to see if I could do it. So I have a headset that does read EEG. It’s made by a company by EMOTIV, and the device that I have is called their Epoc+. I took this headset, and I put it on my husband, and I asked him a series of questions. First, I had him tell me the truth, and then I made him lie to me. What that gave me is a labeled data set of what his EEG looks like when he’s telling the truth, and what his brainwaves look like when he’s lying. From there, it’s fairly easy to feed that data into a machine learning algorithm and build a classifier. So that was one of the projects that I also was working on as kind of a fun thing.

The founder of EMOTIV is a woman named Tan Le and she created device. It’s just a really cool device, and I was able to get a hold of one, and I just thought, “There’s just so many cool machine learning applications, that I have a million more ideas of things I want to do with it.” I’m not a neuroscientist by any stretch, and I do not claim to be one. But I did take one brain science class in college as part of my artificial intelligence work. One of the things that I do remember is that when people tell the truth they typically activate the recall centers in their brain. When they lie, that typically activates the creative centers in their brain. So, I thought, if I have this headset that’s reading from 14 different places on my scalp, might I be able to differentiate between those?

So I did some early research and got some promising results from the first round of testing. Then I’ve worked with several people within our data science team here at Microsoft to expand the work and try to move it forward. But that’s just been a really fun thing to work on. It’s just kind of funny because my first experimenter was my husband, and I had also put it on my manager before too and asked him all kinds of questions like, “Am I going to get a promotion?” He did actually say, “Yes,” and then six months later I did get the promotion, which is pretty funny. So at least in that sense it was true.

insideBIGDATA: Wow, that’s great. I didn’t know husbands can contribute labeled data sets, but you found a way. That’s fascinating.

Jennifer Marsman: Well, I tested with a series of questions that were fairly straightforward yes-or-no questions of, you know, “Where do you live? How many children do we have?” Those sort of things that he couldn’t lie to that well.

insideBIGDATA: That’s great. So on a more philosophical level, what are your thoughts about diversity in the field of data science and how do you see more women become data scientists?

Jennifer Marsman: That’s a great question. Diversity is just so extremely important to any field. I was very lucky to have benefited from a few girls-in-technology camps when I was little. So I benefited from some diversity things very early in my career. But I didn’t really truly get the benefit of diversity until I had my first job out of college with Microsoft. I remember when I got stuck – I used to get stuck on a problem occasionally – I would go talk to these three different people. One of them was someone who had been around Microsoft a really long time, but he came from a Linux background, and he just had a really unique way of thinking. I would go and ask him the question, and he would come up with some answer for me, and I would say, “Thanks.” Then I would go across the hall and two doors down and ask someone else the exact same question. The second person – I don’t even know if he had a degree – but he just had a ton of experience. He had been coding since he was really young, and he had learned a lot from – just learned by doing. He had a lot of real world experience and maybe less college theoretical experience, but just a ton of hands-on experience, and he was just an incredibly smart guy. I would ask him the same question and he would almost always have a different answer than my first guy did. So I would take his answer and say, “Thank you.” Sometimes I’d talk to a third guy who was just out of college and had a lot of the theory, but not necessarily the practical experience and he had a different answer as well.

I found that by doing this I could get the best answers because I could take the short-coming of one solution and apply something that someone else was thinking about. That’s when it really clicked with me that that is the real reason why diversity is important in the industry is because it helps us build better software. There are things like gender and race and things like that that you can use. But also having a good mix of introverts and extroverts on a team. A single mom is going to think very differently about a problem than a guy just out of college. So just taking people from a variety of backgrounds and having them work together really does give the best things. This completely applies to data science as well in what we do and how we use data.

I mean, with data science, the really important thing to remember is garbage in garbage out. So there’s a lot of  questions you have to ask when you’re working with data. What implicit bias existed in the collection of this data? And what can I trust? And what assumptions were made? Because at the end of the day, all machine learning can do is use historical data to make future predictions in the case of supervised machine learning. And so if there’s any kind of biases in your historical data, you’re just going to continue to perpetuate that in your model. So I think having a diverse background and such will help us with identifying these types of things. So definitely extremely important to the field.

insideBIGDATA: Oh, that’s great. We really appreciate those important insights about diversity. I think the audience really appreciates that too. Important stuff. So another question, what’s your take on the recent upswing in interest in AI and deep learning? What do you feel is driving this interest?

Jennifer Marsman: Yeah. I think there’s a couple of things. One thing that’s really driving the interest is probably around the idea of – we’ve made so many advances just in the last 5 to 10 years. I mean, if you look at – in the 15 years since I’ve graduated college, things have changed enormously in the field of data science. Deep learning has fueled a lot of this advancements. So when you look at things like automatic face detection and images and handwriting recognition, and a lot of these other problems like automatic language translation. These types of problems have been solved, are essentially solved problems now, in the last couple of years, using deep learning techniques. I think these things that were considered artificial intelligence 20 years ago, the idea of being able to recognize an emotion or the idea of differentiating between two different people, things that were previously very easy for a two-year-old but very hard for machines are now solved problems.

Because of all of these advancements that we’ve made, I think that kind of has partially refueled and re-surged that interest. The second thing, I think, is around this notion that I’m seeing from Microsoft as well as other companies around the democratization of AI. What I mean by that is taking the power of machine learning and making it accessible to any developer, even people that do not have extensive background in data science and machine learning.

When you think about things like the cognitive services, where you have these amazing pre-built models that can do things like these text analytics jobs and these facial recognition jobs and emotion detection and these types of things, and all I need to do is call a REST endpoint and pass an image or pass some text and get this treasure trove of data back, that’s very empowering because that can realize so many different scenarios that people may not have had access to be able to build before. So I think just kind of making AI available to the people in this democratization of AI in motion, is another huge thing that’s re-surging interest. And then once people see, “Oh, wow. Look at the power. Look at the amazing things I can do,” then they are motivated to want to take the next stop and learn a little bit more about machine learning and how they can actually build their own custom models themselves.

We have a lot of different things available at Microsoft as well. Our Azure Machine Learning product is very popular, which contains these 25 different algorithms to allow people to feed their own data in and kind of wire up the data flow in this really easy drag-and-drop interface that’s available right from a web browser. You can create your own custom models with your own data to solve problems with these well-tested algorithms which is really, really powerful. We also have a deep learning Open Source Toolkit as well, our cognitive toolkit, which we call the CNTK as well. So together you can take kind of the combination of those things that kind of meet your needs, whether you just want to kind of call some pre-built models with the cognitive services, or if you want to go deeper and build your own models to do predictions of, when is this equipment going to fail, or is my team going to make their sales quota, or things like that. You can bring your own data and do something even more powerful with what these tools and technologies enable.

insideBIGDATA: All right. Great. Great points on what’s driving this interest in AI these days. I totally agree. Well, that’s my last question, Jennifer. I’d like to thank you for all the great insights you’ve provided. And I’m glad you could join us today.

Jennifer Marsman: Thank you. It was a pleasure being here.

insideBIGDATA: So for insideBIGDATA, this is the Daniel Gutierrez, thanking you for joining us for another insideBIGDATA podcast. I’ll just leave you with a big “Power to the Data!” Thank you.


Download the MP3


Speak Your Mind