Ira Bell: Today, we talk with Ryan Chynoweth, a Data Scientist at 10th Magnitude. In this discussion, we talk about artificial intelligence (AI) from an architectural, use case, ethics and state of the technology perspective.
Ira Bell: Hi, I’m Ira Bell, CTO at 10th Magnitude, and I have Ryan Chynoweth here with me, a Data Scientist at 10th Magnitude. Ryan, I know you work in 10th Magnitude’s Seattle office, and get to enjoy the beauty of the Pacific Northwest. Some of the pictures we see posted on our internal Slack threads give us the itch to explore the beautiful West Coast. Since we have you today, I thought we’d spend some time talking about artificial intelligence in Azure and then maybe see where the conversation takes us. Does that sound okay?
Ryan Chynoweth: Sure, that sounds great. Thanks for having me. I’m excited to talk AI.
Ira Bell: Great. So, we’ve seen a tremendous amount of growth in Azure, Microsoft’s cloud platform, and specifically, we’ve seen a ton of traction in customers using artificial intelligence in Azure. I was wondering if you could take us through what Azure offers in terms of AI, and maybe describe a generic or typical customer journey as they begin the process of using AI in Azure. I would assume the conversation starts with data, is that correct?
Ryan Chynoweth: Yeah, it always starts with data. AI is powered by data. We use mathematical algorithms to make predictions, augment workflows, infuse that intelligence into applications. So a first step for most organizations is getting what we call AI ready. They have data sources, the data might be all over their organization, in the cloud, on premises, in relational databases, in file shares. It’s everywhere. For a data scientist, this is really hard to aggregate data from all the different sources and join it and kinda manipulate the data the way they need to. So, our first step is getting that data through Azure and into typically an Azure Data Link store.
Ryan Chynoweth: So, once it’s in that store, you have a loose structure to your data, to where you can go to one central location to develop models, iterate, and infuse intelligence. We will typically implement solutions using Azure Databricks to access that data in the data lake. Databricks is an Apache smart platform. It scales extremely well, and really allows you to work with big data, medium data, and even small data. It’s a great platform for development, as well as deploying models. You can deploy streaming solutions extremely easily in Azure. You can slap up at Event Hub, and connect it to Databricks, and you’re making real-time predictions on your data.
Ryan Chynoweth: Once you kind of have that Databricks environment, and you’re developing models, and you’re training it in Azure, it’s time to deploy. And oftentimes we will deploy those models on Azure Databricks using either a batch process or a streaming process that I mentioned before. But as you kind of see, containers and Kubernetes, as that grows in popularity in application development, and even, you see that in AI as well. So things like the Azure machine learning service where you can wrap up your python code and take your machine learning model and put it into a web service that then can scale as needed and handle application predictions or whatever you need, it’s really amazing to see what Azure can do. All the way from storage to deployment.
Ryan Chynoweth: You can then incorporate even Azure DevOps for automated building, testing, and releasing of your models and tracking it that way. So, not only can you deploy your models, but then you can track ’em over time and see how are they performing? Is it getting better? Is it getting worse? And typically we’ll implement these checks to see, hey if it’s getting worse let’s get a data scientist in there to start iterating over this solution again.
Ira Bell: Well that’s, that’s just fantastic. Azure really has a lot to offer, both sort of out of the box, and also in terms of catering to custom and tailored solutions depending on the varying use cases. I guess since we’re on that topic, it’s probably a great idea to talk about some use cases for AI. I’ve personally had a fair amount of exposure to AI in prior roles, and as a CTO at 10th Magnitude, and I’m always amazed when I drop in a meeting at 10th Magnitude and have the chance to see what our engineers are creating. AI is really inspiring to me.
Ira Bell: Actually, my initial exposure to AI, in an actual business case, came about with a company that I’ve been a part of for a while. I co-founded recruit.com with a very brilliant guy by the name of Jeff Nussbaum who came to me and described all of the complex things he wanted the platform to do. So in addition to these complex features he wanted the platform to essentially run itself.
Ira Bell: The platform was built in Azure so it was a natural extension of recruit.com to look at and use Azure for AI. Ultimately we looked at our two biggest pain points for the use of our time which were around images and videos, and jobs or as we call them “opportunities”. So, hundreds of videos and thousands of pictures can be uploaded per day into recruit.com, and it’s extremely important that those images and videos are moderated. So we ended up using the Azure computer vision service which is part of Azure’s cognitive services, and specifically the concept moderator that helps us automatically filter for explicit content, such as nudity, hateful content, acts of violence, etc.
Ira Bell: The other solution we worked with was the Techs Analytics API. This helped us extract proposed key words for our companies and organizations that post jobs, and sort of just suggest them for possible use in their opportunities so that they could import hundreds, thousands of jobs at once.
Ira Bell: I’ve also, kinda switching gears for a bit, heard of solutions in the medical space such as radiographic images being passed through AI services to help detect cancer or other negative health conditions. Doctors and specialists can often be overworked and tired which increases the probability of overlooking something or making a misdiagnosis. I suppose this could be used in veterinary medicine as well.
Ira Bell: So, what have you seen throughout your AI journey, Ryan?
Ryan Chynoweth: Yeah Ira, I just kind of wanted to reiterate the cognitive services are such quick wins for organizations, and you just, a perfect example of what they can do for any organization really. And so that’s exciting that you see it right in front of you. One thing I see organizations really taking advantage of is the custom AI solutions. So, they have their data, they want to use that data to make predictions for their environment, and that’s when they really harness the power of machine learning and deep learning.
Ryan Chynoweth: One example that we have done recently is working for a global restaurant chain where it’s really difficult to understand and predict the demand for a restaurant, let alone thousands and thousands of restaurants, right? So being able to predict the activity, and what’s going to happen in a restaurant at any particular time of day, is really important in order for the restaurant general manager to provide an accurate labor plan. Especially in the US where some states require restaurants to post their shifts two weeks, maybe three weeks in advance.
Ryan Chynoweth: So, what we did is, working on ten plus thousands of restaurants, we wanted to predict for each hour of the day how many people are going to be coming in. And that, in itself, is going to lend an eye towards how many people we need working at a given time. Not only did we do it for each hour of the day, but also each point of sale device in that restaurant. There is a difference between a drive thru and an in-store transaction. So understanding which you need to staff, and how much.
Ryan Chynoweth: Now the cool part to that is, we were extremely accurate in predicting how many people are going to be coming in and ordering food. But what happens if we’re wrong, right? What if we have too many people working? And that can impact, pretty big, on how profitable a restaurant is. If there’s too many people working, or if there’s not enough people working and people are walking in seeing it’s too busy and then leaving. So being able to monitor point-of-sale transactions in real-time, we’re sending that data up into the cloud, and just trying to analyze, “Hey, we have our predictions that we created last night,” or a week ago, whenever we made those predictions and the restaurant GM made that labor plan. “Are our predictions matching up with what’s actually happening right now?” And if it’s not, we need to re-predict for that day, and alert the general manager so then they can then take a preventative action. Maybe they let an employee go early, or they call someone in now instead of two hours later when the peak is going to happen for that day.
Ryan Chynoweth: So it’s really all about, not only employee satisfaction, but also the restaurant profitability as well.
Ryan Chynoweth: Another solution we kind of were working on, it’s a totally different industry, it relates to the legal field. So, I’m sure a lot of our listeners know, lawyer time is very expensive and they track it in very small increments of time. Sometimes it’s as small as six minutes where they say, “For this six minute period I was sending emails about this client.” So something many people might not know is, lawyers are not always very good at associating that time with a particular action. So, on a bill for a legal firm they’ll say, “This lawyer did these time codes,” so they’ll classify their codes. And then also with that, they’ll type in what they did for that time period. So because lawyers are really poor at assigning the correct time code, those bills can come back, they could bounce pretty much where the law firm won’t get paid because the time codes are incorrect. Or as the law firm is getting new cases, new matters, they might want to do some fixed pricing, and so they want to analyze historical cases, historical matters to give their client an estimate on what it might cost them to resolve their problem.
Ryan Chynoweth: So, what we did is a simple deep learning text analysis where we featurized what the lawyer said they did, and we made a prediction on that time code. So, that allows them to, going forward, have accurate time codes, but then also apply their current time codes to historical cases. So they can understand the time breakdown for each matter. So, if they have a new one coming in they can grab an old matter and say, “Oh, it’s probably gonna be pretty similar to this, and it might cost you this much.”
Ryan Chynoweth: Probably one of my favorite use cases is, we worked on it about 18 months ago, it was really cool, it was with a microchip manufacturer. When they make these chips they have images of them. And what they wanted to do was automate the QAing of those chips. And so they had thousands of images, hundreds of thousands of images of chips that were appropriately made, they’re good to go, you can ship ’em. And then they also had images that were detected as defects. And so, we created a convolutional neural network just to simply learn, what does a defect look like and what does a non-defect chip look like? So, as they are producing these chips it’s just a quick prediction saying, “Yep, it looks good.” “Nope, it doesn’t.” It just, simple things like that can allow people who used to have to check chip designs, they no longer had to. You can now take that workforce and apply them to, you know, cooler things. They’re no longer just QAing, they can do something else with their time.
Ira Bell: Thanks for that, Ryan. That’s really neat to learn about all of those varying use cases. I mean, in fact we see a lot of use cases where artificial intelligence is acting more in the proactive space instead of reactive. I suppose the entire purpose is for AI to either take an action or to enable humans to take an action based upon some of the predictions the models have made. And I guess a lot of this is related to the human condition where we have a higher probability of error.
Ira Bell: Speaking of the human condition, one of the biggest conversations happening with AI today is around the ethics of AI. We’ve heard Stephen Hawking and Elon Musk say that artificial intelligence poses the biggest threat to humanity today. Wow! And as we’ve discussed today it can also be extremely beneficial to society, especially if used in cases that ultimately make our lives or businesses better.
Ira Bell: But thinking of the science fiction aspect of AI and ethics, what if one day we are accustomed to seeing a robot as an assistant in every household? What happens, for example, when a robot does something that causes damage to someone else’s property? Who is ultimately liable? And what if that same robot accidentally or intentionally causes harm to another human, or an animal? Is the owner of the robot liable? There are many questions yet to be answered in AI, and the field of ethics in AI is certainly developing.
Ryan Chynoweth: Yeah, it’s a really developing field. To me I think that, it really boils down to what’s the purpose? Is it malicious to begin with? And kind of what are you using in these AI models. What’s behind the scenes that it’s learned from? So, what if models are using race or age to predict an outcome or make a decision. You know, probably not a great idea. And we just, in the end, kind of have to hold these solutions that learn from past data to kind of a similar standard that we hold other humans to.
Ira Bell: Totally agree. For example, what if a healthcare payer’s adjudication system makes a decision on a surgical procedure based upon the probability of whether something’s a legal liability for a company or not? I mean, that sounds pretty terrifying. And hopefully we can mitigate against the probability of bad ethical decisions by artificial intelligence just by kind of a mindfulness, and mentally engaging at the universal risk of this.
Ira Bell: So, in college, I encountered several ethics courses, and I remember specifically learning about rule-based ethics where someone might create a set of rules or some deity might have a set of rules. God, be it. And we’re supposed to follow those, and people could sort of lean upon those. But there are situational ethics as well, such as if you’re under attack is it safe to retaliate, and those sorts of things.
Ira Bell: So, I guess my question for you, Ryan, is… Thinking into the future, can AI actually write code itself yet? Can AI actually look at a problem and say, “You know, I don’t really know what to do here but I’m going to sort of, based upon the knowledge I have, rewrite the outcome”?
Ryan Chynoweth: You know, that is a question we get a lot from a lot of our customers. I’m not a researcher or anything like that in AI field. I implement predictive solutions in the cloud. But it is a fascinating question because I just want to point out, I’ve read an article, it’s called The Unreasonable Effectiveness of Recurrent Neural Networks. It’s about three years old, but I love referencing it because it was one of the first things I read where you see neural networks creating, recreating Shakespeare where it uses the previous word, or the previous couple of words to predict the next one. And when you look at it online, I can’t tell the difference between this neural network writing Shakespeare and actual Shakespeare. Now I don’t read Shakespeare, so maybe someone who actually appreciates that would notice the difference.
Ryan Chynoweth: But you even scroll down through the page and you see mathematical proofs which me, studying math in the university, it’s kinda cool. I mean, it doesn’t make any sense, but just the fact that it can create what looks like proofs. Even sometimes it just says “proof omitted” which a lot of mathematicians will do that. It makes you think that it’s possible. How far away are we from that? I don’t know.
Ryan Chynoweth: So, I say no, AI algorithms can’t write code. But it seems like someday it could be possible.
Ira Bell: Well that’s really neat. I mean, I think of, I don’t have children myself, but I think of probably what a parent might feel like when a three year old child is able to look at a pickup truck and a fire track and still make the association that they’re both trucks when they, to some extent, look completely different. So the associations seem to be really developing in AI. Similarly, when that same child eventually comes to their parent with something entirely profound, and the parent kind of walks away and just thinks, “Wow. You know, I created that.” It will be really neat to see AI doing these things, and sort of going beyond the Shakespeare writing and mathematical models, to see what it really comes up with.
Ira Bell: Thanks for that, Ryan. And so I guess I have to ask, just because I’m so intrigued anytime I meet someone who’s in AI, is what does the current global talent pool look like for AI?
Ryan Chynoweth: Yeah, that’s, I mean you hear a lot that there’s a shortage, that there’s not a ton of people out there. And for the most part, it’s true. There’s not a ton of AI engineers or machine learning engineers, data scientists. It’s growing.
Ryan Chynoweth: I kind of see it in kind of three different spaces. There is the AI researcher. They’re going to be at the big companies researching cutting-edge technology, building self-driving cars, things like that. There’s gonna be someone who can use data, train a machine learning model, and put it into production. A more enterprise type data scientist. And then there’s gonna be kinda going back to someone who’s gonna use like the Microsoft cognitive services. That’s not really someone who I would classify as a data scientist, a machine learning engineer, AI engineer, AI researcher. They’re more of a software developer who can then consume predictions, very intelligent predictions, into their applications. So there is those three spaces, and overall I would think that that enterprise data scientist is hard to find. A lot of ’em will be very specialized in training machine learning models, but they’ll need assistance getting that model into production.
Ryan Chynoweth: So it’s difficult, but it’s not hopeless. There is a pool. It’s not big though.
Ira Bell: Well that’s great. Well I’m certainly glad you’re part of our pool, I can tell you that. So let me ask this, where would someone even go to get started in the field of AI if they’re interested? Where does someone even begin?
Ryan Chynoweth: Yeah, there’s a ton of online courses out there. I would just, yeah, just google, go to google and search “data science courses”. Most of ’em are really great. And then in addition to that, I love the data science website Kaggle, kaggle.com. They host data science competitions. I’m always constantly doing one competition on the side just to make sure I stay up to date, and it’s just fun for me. So I would kind of point to both of those. Online courses are great, and those competitions allow you to get applicable use cases and solutions.
Ryan Chynoweth: Even if you are doing the Kaggle competitions, I would totally recommend getting a public repository out there and showing people “Hey, check out this competition I did.” Now, the competitions are extremely competitive. I know a lot of times when I enter ’em I’m in the top 3,000, but the difference between place 3,000 and first place is really not that big. A lot of people enter. It’s within five degrees of accuracy or something like that.
Ryan Chynoweth: But yeah, online courses and data science competitions are my favorite.
Ira Bell: Well thanks for that. Well, with that I’d like to thank you very much for your time, Ryan. It’s definitely a pleasure to work with you, and I hope that we can have you as a guest on our podcast again soon.
Ryan Chynoweth: Yeah, thank you for having me.
Ira Bell: Thanks for listening to The Art Of Digital Disruption. At 10th Magnitude we’re proud to create the path for organizations to stay competitive and disrupt their industries. And for more information on innovation, and how you can disrupt your industry, visit 10thmagnitude.com/agilityquadrant/ and download our latest whitepaper. You can also look at the format for our AI Ideation Workshop by visiting 10thmagnitude.com/aiideationworkshop/. Thanks for listening.