The conversation this week is with Blair Newman. Blair is an accomplished executive with more than 25 years of experience creating business value within the information technology industry. He's a decisive leader managing and overseeing PnLs, building pipelines, and generating multimillion-dollar cost savings by acquisitions and IT transformations. He is currently the Chief Technology Officer at Newton AI where he and the team helped companies automatically build extremely tiny and explainable models without needing a Machine Learning background. Prior to his current role. He was head of computing services and solutions at T-Systems North America for more than 10 years.
If you are interested in learning about how AI is being applied across multiple industries, be sure to join us at a future AppliedAI Monthly meetup and help support us so we can make future Emerging Technologies North non-profit events!
Resources and Topics Mentioned in this Episode
Blair Newman 0:00
Now don't get me wrong. I'm not saying we don't need mechanics. But I need a vehicle just really primarily to get me from point A to point B. And as an embedded engineer, oftentimes you said, you know, I don't want to be a data scientist, I can't be a data scientist. But I need a way to get from point A to point B so that the devices that I'm working with, they can now bring additional value to my customers. And so for me, the use cases that I really enjoy is not so much a technical one so much is when I have an opportunity to sit down with a partner sit down with an end customer at times and show them and illustrate how they can get from point A to point B and to see that light bulb and to see that value transition from technology to gratification. Those are the use cases that I that I truly enjoy.
AI Announcer 0:51
Welcome to the conversations on applied AI podcast where Justin grumman's and the team at emerging technologies know of talk with experts in the fields of artificial intelligence and deep learning. In each episode, we cut through the hype and dive into how these technologies are being applied to real world problems today. We hope that you find this episode educational and applicable to your industry and connect with us to learn more about our organization at applied ai.mn. Enjoy.
Justin Grammens 1:21
Welcome everyone to the conversations on applied AI Podcast. Today we're talking with Blair Newman. Blair is an accomplished executive with more than 25 years of experience creating business value within the information technology industry. He's a decisive leader managing and overseeing PNLs, building pipelines and generating multimillion dollar cost savings by acquisitions and it transformations. He is currently the Chief Technology Officer at Newton AI where him and the team helped companies automatically build extremely tiny and explainable models without needing machine learning background. Prior to his current role. He was head of computing services and solutions at T systems North America for more than 10 years. Thanks, Blair for being on the program today.
Blair Newman 1:59
Thanks, Justin for having me on. Appreciate.
Justin Grammens 2:01
Well, awesome. I mean, I gave a little bit of a background for you here on the intro. How long have you been the CTO where you're currently at? And how did you end up landing in that position? I guess.
Blair Newman 2:12
Okay. Yeah. So I've been with the organization for now, it will be seven years and probably a little over a month. So I've had the opportunity to see the organization grow, until, to some degree, we kind of went through our old transformation throughout this period. It's been a fun ride.
Justin Grammens 2:31
Awesome. Well, so I'll biggest Newton?
Blair Newman 2:35
I would say it's, it's our brand. It is the brand of our tiny ML platform, the organization that I work for Bell integrator, we've been around for a little over 17 years. And as I've mentioned myself, I've been with the organization for going on seven years. So we've historically been a systems integrator. And as I mentioned, one of the things that we were looking to accomplish, since I had the opportunity of joining the organization, is my approach in my philosophy has always been, you know, how do I bring, let's say, technology? Or how do I bring services to the fingertips of your customer, whether it's our direct customer, or if we're working with a partner, let's say they're in customer. And that's been one of our core objectives, since I've joined the organization. And I think, as we have an opportunity to talk a little bit more about what we're doing with Newton, hopefully, I have an opportunity to illustrate this a little bit more.
Justin Grammens 3:31
Well, we'll spend a lot of time talking about as much, much as you would like to talk about Newton. One of the things I do like to talk to people about I mean, since you're sort of a thought leader, I think in this in this space of artificial intelligence and machine learning tiny ml, in fact, I saw you present at the 20 ML Summit. That was awesome. I really, really enjoyed the presentation. How do you define AI? And like machine learning? If somebody isn't outside of the field? How do you tell people what you do or what your product does at more of a layman's term level,
Blair Newman 3:58
I like to explain it hopefully, as simply as possible, I've actually had the opportunity to establish a AI club at my youngest son's High School, some of the participants of some of his peers, they were new to us at any level of technology, some have began to get a little bit of an exposure to artificial intelligence. But one of the analogies that I used that I believe that was able to resonate with them and hopefully also for those listening is that when you begin to think about machine learning to a large degree, you can really contrast this to as an individual, your path to learn and starting off with normally how do we educate ourselves, right? We normally educate ourselves with some historical information where oftentimes this information is presented to us maybe in the form of a book. Maybe nowadays, were reading some various white papers. So you need to go through that process of Education, right are learning to a large degree, right. And then once you kind of go through that process of understanding, let's say some of that background, then at that point, you'll be able to hopefully make some decisions based upon some of the historical information that you may have picked up along the way. And when you begin to think about machine learning, what we're trying to accomplish is we're trying to replicate some of that decision making that we make our we've made, as we've gotten older, more experienced, etc. So you typically want to start off with some historical information, hopefully, factual information is still the old tentative junk and junk out. Still, the old rules the day. So you want to make sure that you have some information that you can then leverage and use as it relates to training the model are creating the model. And just like when you're in school, you read chapters one through five, and then you have an instructor and they say, Hey, you know, I'm going to test how well you've retained that knowledge. And if whether or not you can answer questions correctly. And we do something very similar. When it comes to machine learning or artificial intelligence, we want to validate, we want to test the model to ensure that it is making predictions with as high or higher level of accuracy as we possibly can get. And as we grow, we continue to learn and sometimes this turns into an iterative cycle. And the same thing with machine learning over time, maybe your model may begin to decay. And you may need to go through you may need to give it some additional information so that it can either you may be build a new model and start from scratch. Or you may do something what we call reinforcement learning where you're training the model on the fly. And then you're hopeful that the predictions that you're making will continue to be as accurate and valuable as possible.
Justin Grammens 6:53
I love that analogy. And I was just thinking about kind of going back to school for a quote unquote, tune up, I guess, right? If you if you're there's other classes and coursework you need to take, then yeah, that's just kind of getting your model improvements over time making it smarter and smarter. This is really cool that they're kind of embracing this technology and sort of like interested in playing around with it do they? Do they see the tie between machine learning and artificial intelligence and say, like Google, or Siri or, or Alexa or any of that type of stuff,
Blair Newman 7:21
they definitely know that it's out there, they definitely know that they interact with it daily, I think probably one of the most eye opening experiences for them. And it's an ongoing experience for myself, as well as how actually invasive machine learning is in our lives today. So one of the things that I had an opportunity to illustrate was just just played a video, and I just asked him to count how many times in that video did this particular individual interface with some form of either artificial intelligence or machine learning, etc. So that was quite an eye opening exercise for them to begin to think expansively as to really how how often you are interacting with some level of technology that oftentimes at least for themselves, they didn't know was influencing, you know, their actions, their behaviors, their thought processes, etc. So that was definitely an eye opening experience for them.
Justin Grammens 8:20
Yeah, for sure. What's a day in the life of a person in your role at your organization?
Blair Newman 8:26
Well, one of the tenants that I mentioned when I started with the organization is, how do I bring these services to the fingertips of our customers. And when we began the journey of building out this platform, which was roughly around, let's just say, going on about four years ago, we started out really with one mission, how do we enable machine learning for everyone? And we meant that in a very literal sense, how do we make machine learning available for everything. And now as tiny ml was kind of matured, and really began to grow exponentially, I would say, when you think about edge devices, when you think about bringing intelligence to the edge, is really kind of really pulling forward that vision that we had at the time. Now for myself, when when I express myself, when I think about a day in the life, one of the things that I truly enjoy is the say, having the opportunity to turn the light bulb on for anyone. A lot of times I find myself both internally, as well as externally, communicating with potential customers and partners, to really help them understand and oftentimes just articulating how to say, embedding a model can really bring intelligence to the edge. Because, you know, to some degree, we've almost come full circle. Maybe 10 years ago, there was this huge, huge movement to bring everything to the cloud. It was to some degree it was was a race, you know, how much can you put into a cloud, whether it was public or private, but it was this huge, huge race to put as much as possible in the cloud. We've almost come full circle. And we're saying, how much now can we bring to the edge? How much intelligence can we bring to the edge? When you challenge the organization now must shift a little bit internally, when you challenge the organization and you say, I want to make machine learning, or I want to make tiny ml available to everyone. The first response sometimes is, well, that's not possible. But you know, those are some of the headwinds that we had to address back then. And to some degree, those are the headwinds that we have to address now. Whether it is just simply just building a model automatically from a zero code perspective, whether it's producing code, and I say, hey, I want to, I want us to be able to produce a model that is device agnostic, I want us to be able to produce a model that will allow someone who has no educational experience as it relates to being a data scientist, to be able to successfully implement a model and embed that model into their MCU in two days. So my objective is to keep pushing what sometimes may seem as the impossible and that that includes interfacing with customers, that includes interfacing with internally, and I'll give you another example. And I'm sure that at the tiny ml event, you probably heard me speak to this, I would probably say even still today, 99% of all models that are built, they start off with a fully connected network, they're extremely large. And then there are a number of different tried and true technologies, that everybody is applying quantization, pruning excetera to reduce the size of those models, so that they can be embedded into memory constrained devices. And no matter where you go, you even go now to Google or STM micro. And they're producing, they're building on the same approach to same methodology. And what we said, which hopefully now can bring to home a little bit, when I say My objective was to make them impossible reality, where we build all of our models, neuron by neuron from the bottom up. And we're one of the only organizations that take this approach, which allows for us to now get to the point where we began to set the standard, that all of our predictive models, all of our machine learning models that we produce are less than a kilobyte in size as crazy. And if there's anything larger than that, then we have to have a discussion. So we've really kind of taken the bull by the horns in an attempt to kind of redefine how machine learning is approached, I vacillate back and forth between helping customers to understand how this is even possible and how they can get value out of it, as well as continue to impress upon individuals in our organization to keep pushing it
Justin Grammens 13:14
Sounds like a fun role for sure I as you were talking about your product, which I want to have, you know, maybe you explain a little bit more, maybe some some specifics, or some customers you work with, or, you know, whatever, I have you heard of teachable machine. It's a project, I think it was put together by Google, but it's, it's an in browser solution. And basically, you can turn on your camera and you can have yourself smile 50 times and then frown 50 times, or whatever it is. And then so you're essentially trained the model. And the only thing that I think it's really cool about it is, number one is you can you can generate a TensorFlow model, you can download and do and do it wherever you want. But also just the getting to your point of, you know, putting the power in anyone's hands, the accessibility, you know, a basically a 10 year old could train a model through this through this tool. And I love that concept. Because I'm a technologist at heart, I actually teach at a local university here as well. So I feel like I'm, I'm definitely an educator. And the more the more, I guess, knowledge you put in the more hands on, you know, throughout the world, and, and even younger age groups, sort of like the better. So I really commend you guys on what you are trying to do. The fact that you can do stuff in 1k. Two is is phenomenal. I'm assuming there's a lot of like, intellectual property built in. Have you guys been filing patents and stuff like that in that area?
Blair Newman 14:28
Yeah, absolutely. We, as I mentioned, we began to build out our AR platform little over four years ago. And you might say that initially, just for the sake of today's discussion, we were just doing auto ml. So we worked in a number of different verticals, let's say in the retail space in the financial services space. And our platform allows for customers to do more than just tiny ml. They can do predictions via the web, you can publish your models via an API, etc. It's actually kind of funny, we actually kind of stumbled across producing tiny models. It wasn't our objective, our objective was that we worked on a project, it was with a local auto retailer. And they wanted to produce a application, a mobile application that allow for their customers to take a picture of the vehicle. And then it would identify that vehicle, and then map the accessories that were associated with that vehicle. Because I'm sure if you know, if you go into AutoZone, or Pep Boys and you want to get an accessory, you got to pull out the big manual, you have to find your particular model of the car, you have to find a part number. And you know, it's very laborious to take that approach. And he said, you know, we don't want to do that. Just, we want to give them the mobile phone, take a picture, identify the accessories, and then we're also looking to extend our reach outside of their brick and mortar location. And they said, hey, just go back. 10 years, I've been identifying vehicles, we probably had the time to build the models, well over 5 million images to make it work. But the models were huge. They were very large. We did finally get it to work. But it was a clunker. It was a it was a clunker we got it to work. But it was a clunker. So one of our objectives was, hey, how can we build models but not requiring, you know, such a significant amount of data. And in the end, it just wasn't it wasn't scalable. And that's where we kind of challenged each other internally to say, Okay, let's do something different. And that's what we flipped the paradigm on his head, just eliminated starting out with a fully connected network. We don't use TensorFlow is an algorithm and framework that we've developed completely in house and it is patented. And as a result, even though we were just serving, let's just say some of the major verticals, one of the things we began to realize were Wow, these bottles are really small. As tiny ml picked up, we said, oh, wow, we can really we're onto something here. We think we can play in this area. And so it wasn't actually our intention to really be active in this space. But this is more recently where we kind of almost function exclusively in the tiny ml space, since we've had an opportunity to be so successful in this area.
Justin Grammens 17:19
That's awesome. You've been sort of talking through it, I guess what you guys do, but I mean, you guys offered this platform. Right? And I can, and you mentioned API, so maybe you can talk me through I guess how you guys work? This is this. I mean, I need to have an MCU. I do I need to run some sort of operating system on my local system. And how do I bring your software into what I'm building,
Blair Newman 17:39
I'll take you through our guests our key three tenets as it relates to being able to really make AI available to everyone. So starting number one with ease of use, of course, there's a number of platforms that, you know, are auto ml based, etc. But one of the things that we really focused on is having an automated pipeline, really a three step pipeline, that all of our customers go through, as I've mentioned, is zero code 100% zero code, as it relates to being able to produce one of the machine learning models. So our customers, they drag and drop their data onto our platform, we kind of focus in a couple of key areas regression classification, where there was binary multi classification. We also work with time series audio. So digital signal processing, also included in our platform is NLP. Now, one of the areas where we have not integrated into our platform at this time, from an auto ml perspective, is video and imaging. But all of the other task types or use cases, they're available where a customer, you just drag and drop your data, you identify your target variable. So depending upon which sensor you're using, which MCU or what just what your use case is. And depending upon your task type, and your device, then you would select if you're using a 32 bit device, if you're using floats, etc. One of the next areas that we really, I would say we kind of kicked the door down is that our models, in addition to being device agnostic, also natively come out. And they support eight bit and 16 bit microcontrollers. So this also created another opportunity for our partners, our customers to really extend the reach of intelligence. So you will then select if whether or not you have an a big controller, 16 bit, controller, 32 bit controller, etc. Then after that, you just click Start training, and then we automate everything else from there. Now, once the training is complete, you go into the third step of the pipeline. Now if you want at this time, you can run predictions via the web. So say just for some reason, you haven't gotten to the point of embedding you just want to you have a holdout data set that you wanted to As validate the accuracy, so that's available to our customers. We also then at that time, we will publish a REST API, which if you wanted to run predictions remotely, and query the API, you can, you can do that. And then lastly, of course, from a tiny ml perspective, then our customers can select and hit download. And then at that point, they can download the model, which comes in C code. So at that point, when it's downloaded, then they can then move to the next step of embedding that model into their device of choice. So even though it is our own proprietary algorithm, all of the models that we produce, they're open, so they can be converted into onyx, h five, format, etc. So there's no limitations in regards to being able to take advantage of the models that were producing. So that's what I would say is purely, let's say, from an easy use perspective, as well as from whether you want to say as portability, or the ability to access, really any device that you may be working with for your particular use case. And then lastly, we began to talk about making AI or machine learning available to everyone, our platform is available 100% free for anyone. So you get full use of our platform, where you can build your models, you can test your models, and you can download the models with no tie backs to our platform whatsoever. And that's available to anyone, irrespective of the use case that you may be working with. And we work very actively with, with a number of different whether they're just doing it from educational purposes, or whatever the case may be. So those are kind of the three areas that we really kind of touch on is from an expense perspective is 100%. Free. And then of course, we've kind of knocked down all of the headwinds as it relates to usability.
Justin Grammens 22:01
That's fabulous. Yeah, I mean, on your website, you can basically start for free. Is there a cap, I guess, on on so many models to train? Or when do things have to flip over and start costing you money?
Blair Newman 22:13
No, absolutely not. So unlimited training, unlimited predictions, our business model is that we typically are working with various partners. So various silicone providers, they are looking to extend their value proposition in their service chain, to enable their customers to come directly to them, and then not only be able to purchase the silicone or the sensor, but then be able to go through the entire service chain, from purchasing the silicon to then having an embedded model. So we will work with our partners to help define and establish use cases that their customers can consume. So obviously, it drives a lot of value for a number of our partners, because now their customers can come to them, not only to just purchase the silicon, but also to go through the entire use case and implementation. So we'll be shoulder to shoulder, sometimes underneath the umbrella of some of our partners, enabling their community to go through and implement and embed models into their respective devices. So we will have a commercial engagement with a number of those partners on a use case by use case basis, where they identify where their community can get the most value out of
Justin Grammens 23:35
awesome. Yeah, I just looked into there's tons of use cases you have here, you know, edge devices for home monitoring, there's home security, automation, pet tracking, yeah, various case studies, do you have a favorite one, I guess, have you guys have done recently,
Blair Newman 23:48
we do make available a number of what we call pre trained use cases. So the data set is there, we provide all the necessary information in order to let's say, produce the model, I would say a couple of the use cases that I really enjoy is really anomaly detection. And to some degree, the say, fault prediction. I enjoy these because especially anomaly detection, because oftentimes those can be some of the more challenging use cases to implement. But one of the things that I did highlight to you is that my passion is behind, bringing the light bulb on, and so have an opportunity to work, let's say with the embedded community or just different members of that side of the house, because in reality, there's a finite amount of data scientists on the other end, especially when you know you're working in a silicon space, there's numerous embedded engineers that oftentimes don't know where to begin, have some questions and doubts about, Hey, can I do this? Is it even possible and I don't have the expertise. One of the things that I like to tell them one of my own favorite personal analogies is, at least for myself, I live in California, I live in the Bay Area. And I have to drive. Unlike New York where maybe you can to a large degree, maybe just rely on the public transportation or the subway. But for myself, I have to drive. And I have to get from point A to point B and getting from point A to point B to me is critical. Now, that means more often than not, I need a vehicle, I need a car. But you know what, I don't want to be a mechanic. Now, don't get me wrong. I'm not saying we don't need mechanics. But I need a vehicle just really primarily to get me from point A to point B. And as an embedded engineer, oftentimes you said, you know, I don't want to be a data scientist, I can't be a data scientist. But I need a way to get from point A to point B so that the devices that I'm working with, they can now bring additional value to my customers. And so for me, the use cases that I really enjoy is not so much a technical one so much is when I have an opportunity to sit down with a partner sit down with an end customer at times and show them and illustrate how they can get from point A to point B and to see that light bulb. And to see that value transition from technology to gratification. Those are the use cases that I that I truly enjoy.
Justin Grammens 26:16
Awesome. My undergrad was in applied math. And I always tell people that I didn't like to just solve for x, right? I actually want the applications of math actually applying in the real world. And so those were the things I always thought was really cool, you know, solving an equation what what do you do? But does that tell me how far I threw this football, for example? Well, now there's something there. Right, really good philosophy. And I think one of the things that I'd like to ask people that are have been on the program, kind of as we get near the end here is, you know, people entering the field, how do you would advise them? Or how would you advise somebody that wants to get into this space? Like, are there classes, books, open source communities? I mean, obviously, probably, downloading Newton and just start playing around with it is something cool, but yeah, how would you advise somebody who wants to get into this AI ml space?
Blair Newman 27:02
Yeah, there's a number of different OnRamps that Harvard, they partnered with Google, they have an excellent program is so many on ramps, it's amazing. But what I would say is, maybe to some degree, historically, and even still, today, you still kind of run across this belief system, that in order to do effectively, and we're talking about driving value, in order to do machine learning, you have to be an expert in math, or you have to be a hardcore coder that sits in a corner in the cubicle, you know, late at night, pounding away on the on the keyboard. But for me, the reality is that I think the real area that I would encourage anyone to focus on is being a domain expert in the area of where you're looking to focus. Understanding the data is by and far the most essential step that you need to understand and take, when you're building a model, irrespective of the tool that you're using. If you decide that you know, you want to go through, you want to build a model manually. And you know, you want to use our you want to use Python, or whatever your tool of choice, the most critical and essential step is understanding the data in the domain that you're working at. If you struggle with this, then the possibility and the probability of you being successful, it really takes a significant nosedive earlier, we kind of just kind of joked around that, hey, you know, junk and junk outs, deal rules today, even after all of these various years. But even more importantly, understanding the features or the data points that you need as inputs to your model is absolutely critical. That's a part of our secret sauce. Whereas you may have existing frameworks that fully connect to every feature. And then you go through and you, you know, with a scalpel, and you disconnect, different features that are not as important. That's the process, let's say of pruning as an example, where our focus is we only connect to selected features that have the highest probability, or highest predictive power. And a part of that is having that domain expertise to know which features are really going to contribute to what you're looking to predict. So that's what I would really encourage people to start and really kind of focus their attention on. Interesting.
Justin Grammens 29:38
Yeah, I mean, at the end of the day, it's it's how are you going to change? How are you going to use the power of AI and machine learning, I guess, to actually change the way your businesses run or bring value, I guess the market and as you were talking, you know, it feels like Newton would speed up that process that will allow you to do something a lot more iteratively rather than if you were just have to hire a bunch of data. Scientists go through this whole rigmarole then just to find out at the end of the day, what we should we should have brought in this data instead. Right? Can you guys I feels like there's some value there, it's, there's always, you're gonna move faster, the faster you can iterate.
Blair Newman 30:13
Yeah, I mean, building a model is resource intensive. And you can say, resource from a data perspective from a resource, or whether it's, let's say, you need to hire a number of different experts. And for sure, building a model was an extremely iterative process just to produce a model, let alone it's whether or not it is bringing your organization value. So we definitely assist in accelerating time to market. So if you are just kind of entering into this particular space, or if you've kind of reached a space where, hey, you kind of got it right now you're looking to scale. Right? So there's the right tool, obviously, for every situation, but in this particular case, yeah, absolutely. We definitely missed one of our key tenets is, you know, being able to accelerate your time to market go through your test out your hypotheses fairly quickly. And to really, really draw down or eliminate as much of you know, those headwinds when it comes to the number of resources that you would need, whether it's technically or whether it is from a human perspective or a resource perspective, as well. So we we really assist in both areas.
Justin Grammens 31:22
Yeah, for sure. You know, in all these podcasts, I always have liner notes in there. So I will put all information, of course, links to your websites and everything that we talked about, how do people best reach out to you Blair,
Blair Newman 31:33
my most active platform is LinkedIn. So I can always be found on LinkedIn. And I'm also not shy in sharing my email address. So first, I last had newton.ai Newton that AI can also be reached as well. So sure,
Justin Grammens 31:51
cool. Well, I liked it, you're giving back to the local school, I guess your core values with regard to like, you know, what, what the company is trying to do, I think is very, very inspiring. Is there anything else that maybe I hadn't brought up that you wanted to highlight or talk about it? I think we've talked about most of the areas that we wanted to do today?
Blair Newman 32:11
Well, I think I think we've done an excellent job in a short period of time that we that we that we do have. And again, I want to extend my thanks for inviting me. I certainly appreciate it. Hopefully we'll we'll do it again. Or maybe we'll have an opportunity to do something together at an upcoming event.
Justin Grammens 32:28
Yeah, for sure. For sure. Blur. All right. Well, thanks again for your time. And yeah, definitely. We'll be in touch and see you around in the AI community. All right, thank you.
AI Announcer 32:39
You've listened to another episode of the conversations on applied AI podcast. We hope you are eager to learn more about applying artificial intelligence and deep learning within your organization. You can visit us at applied ai.mn To keep up to date on our events and connect with our amazing community. Please don't hesitate to reach out to Justin at applied ai.mn If you are interested in participating in a future episode. Thank you for listening