Understanding the landscape of AI techniques

Jared Lander, the organizer of NYHackR and general data science guru, joined us to talk about the landscape of AI techniques, how deep learning fits into that landscape, and why you might consider using R for ML/AI. Sponsors: Hired https://hired.com/practicalai – Salary and benefits upfront? Yes please. Our listeners get a double hiring bonus of $600! Or, refer a friend and get a check for $1,337 when they accept a job. On Hired companies send you offers with salary, benefits, and even equity upfront. You are in full control of the process. Learn more at hired.com/practicalai https://hired.com/practicalai. Rollbar https://rollbar.com/changelog – We catch our errors before our users do because of Rollbar. Resolve errors in minutes, and deploy your code with confidence. Learn more at rollbar.com/changelog https://rollbar.com/changelog. Fastly https://www.fastly.com/?utm_source=changelog&utm_medium=podcast&utm_campaign=changelog-sponsorship – Our bandwidth partner. Fastly powers fast, secure, and scalable digital experiences. Move beyond your content delivery network to their powerful edge cloud platform. Learn more at fastly.com https://www.fastly.com/?utm_source=changelog&utm_medium=podcast&utm_campaign=changelog-sponsorship. Linode https://linode.com/changelog – Our cloud server of choice. Deploy a fast, efficient, native SSD cloud server for only $5/month. Get 4 months free using the code changelog2018. Start your server - head to linode.com/changelog https://linode.com/changelogFeaturing: Jared Lander – Website https://www.jaredlander.com/, GitHub https://github.com/jaredlander, X https://x.com/jaredlanderChris Benson – Website https://chrisbenson.com, GitHub https://github.com/chrisbenson, LinkedIn https://www.linkedin.com/in/chrisbenson, X https://x.com/chrisbensonDaniel Whitenack – Website https://www.datadan.io/, GitHub https://github.com/dwhitena, X https://x.com/dwhitenaShow Notes: Lander Analytics https://www.landeranalytics.com/New York R Conference https://www.rstats.nyc/New York Open Statistical Programming Meetup https://nyhackr.org/Books “R for Everyone” by Jared Lander https://www.amazon.com/dp/013454692X“Applied Predictive Modeling” by Max Kuhn and Kjell Johnson https://www.amazon.com/dp/1461468485“Deep Learning with R” by Francois Chollet and J. J. Allaire https://www.manning.com/books/deep-learning-with-r-second-editionUpcoming Events: Register for upcoming webinars here https://practicalai.fm/webinars!

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Baiting for change log is provided by Fastly. Learn more at fastly.com. We move fast and fix things here at change law because of Rollbar. Check them out at rollbar.com and we're hosted on Linode servers. Head to linode.com slash change log. This episode of Practical AI is brought to you by hired. One thing people hate doing is searching for a new job. It's so painful to search through open positions on every job board under the sun. The process to find a new job is such a mess. If only there was an easier way. Well, I'm here to tell you there is. Our friends at hired have made it so that companies send you offers with salary, benefits, and even equity upfront. All you have to do is answer a few questions to showcase who you are and what type of job you're looking for.

• 0:45Speaker 00

  They work with more than 6,000 companies from startups to large publicly traded companies in 14 major tech hubs in North America and Europe. You get to see all of your interview requests. You can accept, reject, or make changes to their offer even before you talk with anyone. And it's totally free. This isn't going to cost you anything. It's not like you have to go there and spend money to get this opportunity. And if you get a job through hire, they're even going to give you a bonus. Normally it's $300, but because you're a listener of Practical AI, it's $600 instead. Even if you're not looking for a job, you can refer a friend and hire will send you a check for $1,337 when they accept the job. As you can see, hired makes it too easy. Get started at hired.com slash Practical AI.

• 1:38Speaker 00

  Welcome to Practical AI, a weekly podcast about making artificial intelligence practical, productive, and accessible to everyone. This is where conversations around AI, machine learning, and data science happen. Join the community and slack with us around various topics of the show at change.com slash community. Follow us on Twitter, we're at Practical AI FM. And now onto the show.

• 2:03Daniel Whitenack?

  Well, this is Daniel Whitenack, your data scientist co-host, and I'm joined by Chris Benson, who is the esteemed AI strategist specializing in deep learning. Chris, how are you doing?

• 2:17Chris Benson?

  I'm doing great today. How are you, Daniel?

• 2:19Daniel Whitenack?

  Doing great. And today, we have a very special guest I'm excited about. So tell me, Chris, if you opened up your editor, whatever that might be, what language would you start programming in?

• 2:34Chris Benson?

  Well, usually, it's going to be either Python or Go for me most of the time. But I understand we're going to do a different language today, which I'm looking forward to.

• 2:43Daniel Whitenack?

  We're going in a different direction, actually. Well, I started out, you know, back in the day doing FORTRAN and then kind of moved into the Python world eventually to Go. But then, every once in a while, I kind of dabble in this other language called R. And I'm very happy that we have one of the great people from that community, the R language community with us. We have Jared Lander. Welcome.

• 3:12Speaker 03

  Hi, folks. It's wonderful to be here.

• 3:15Chris Benson?

  Hey, thanks for coming.

• 3:17Daniel Whitenack?

  Yeah.

• 3:18Speaker 03

  Thank you for having me.

• 3:20Daniel Whitenack?

  So Jared, tell us, give us a little bit of background about yourself. Where are you coming from and what do you do?

• 3:27Speaker 03

  So I am what I would like to call, depending on the audience, either a statistician, a data scientist, I do machine learning or AI, depending on who's listening.

• 3:38Daniel Whitenack?

  And that also depends on if you're on a plane and if you actually want to talk to the person next to you, then probably you say AI. And if not, then you say statistician. Is that how that works?

• 3:48Speaker 03

  That's pretty accurate.

• 3:50Chris Benson?

  Hey, I think I did meet you before.

• 3:55Daniel Whitenack?

  Awesome. Well, what are you working on right now? And I understand you have your own business, right?

• 4:03Speaker 03

  Yes, I do. It's Lander Analytics. We are a data science consulting, training and events company. So at any given moment, we have many different things happening, whether it's client projects, we're teaching people how to use R, or putting on events.

• 4:19Daniel Whitenack?

  Awesome. Yeah. And actually, is it one of those events that I met you? And it was a great event. It was the NYC R Conference, or the R Conference NYC. Which one of those is right? I forget.

• 4:33Speaker 03

  We call it NYR.

• 4:35Daniel Whitenack?

  NYR. Ah, even better. Yeah. No, this is a really great event. I have to say, not only was the content amazing, the technical content, but the catering was the best out of any conference I've been to. So, round of applause on that.

• 4:53Speaker 03

  Nice. Thank you. We pride ourselves on the food. This year, we stepped it up a notch. We had baristas on site. We had ice cream sandwiches, pizza delivered in waves. We really try to make it a party more than a conference.

• 5:06Chris Benson?

  You mean the food doesn't all have to start with R?

• 5:08Speaker 03

  Yeah, that's a fair one. Maybe we could try next year, find, you know, round pizzas only.

• 5:15Daniel Whitenack?

  Well, I mean, there is an impressive number of puns in the R community, as I've learned.

• 5:21Speaker 03

  Yes.

• 5:22Daniel Whitenack?

  That seems to be a point of pride as well.

• 5:26Speaker 03

  Yes, they are. We are, and not only puns, but people are to double puns. They really, they really go deep in how much they pun things.

• 5:36Daniel Whitenack?

  Yeah. So maybe, maybe we'll just devote a show to data science puns one of these days, and we can, we can have you back on to discuss that.

• 5:47Speaker 03

  That would be great.

• 5:48Daniel Whitenack?

  Yeah. I'm looking forward to that show. Yeah, I am too. But also, so tell us a little bit. So you're involved in the conference, the New York R Conference. You also are really involved in the meetup scene in New York, I know. I'm actually, even though I'm not in New York, I'm part of the Slack that you guys have going there. It's very, very active. And I just appreciate, I think, how much, you know, the community there. Even though I'm not a, like, really heavy R person, was very welcoming to me when I was there. And I think in general, it's just a really welcoming and awesome, awesome community for newcomers to data science and machine learning and AI.

• 6:30Daniel Whitenack?

  Why don't you tell us a little bit more about that community and what you guys do and how that came about?

• 6:35Speaker 03

  Absolutely. The, one of the hallmarks of the R community in general is its friendliness. And not just, you know, "Hey, how you doing?", but it's welcoming environment that really tries to, it really strives to make everyone who walks in the door, either physically or virtually, feel welcome and happy and like you're a part of the community. And I do think that is one of the strengths of the R language, is this community behind it. So the meetup was actually founded back in 2009 by Josh Riech. And shortly thereafter, by the second meetup, and it was a monthly meetup, Drew Conway had taken over. And every good scientist knows Drew Conway.

• 7:17Speaker 03

  He's one of the luminaries, he's one of the original, I'm gonna call him the old guard of data science. And after running it for about two years and growing it to 1,800 members, Drew asked me to take over the meetup. I actually took over at the time with Wes McKinney, who's technically my co-organizer. The New York R Meetup, which is now called the Open Statistical Programming Meetup, so that it is welcome to all languages. And indeed, we've had Python, we've had Go, we've had SQL. We've done a number of different languages, as long as they're open source. It is primarily focused on R, but we do allow R in Friends. Same with the conference.

• 7:58Speaker 03

  It started in a room at NYU back in 2009 founded by Josh Riech and was quickly organized by Drew Conway. And everyone knows him, he's one of the old guards at Data Science. So after about two years of running it and growing it to about 1,800 members, he turned over the reins and I am very proud to say we are now about to cross 10,000 members.

• 8:24Daniel Whitenack?

  That's crazy. Wow.

• 8:26Speaker 03

  Yeah.

• 8:27Daniel Whitenack?

  So how many, how many of those 10,000 show up at a, at an event? I know like the proportion, you know, on meetups is, is, uh, is smaller, but you still must have a lot at the in-person events.

• 8:40Speaker 03

  We do, and since we're in New York, space is our limiting factor.

• 8:45Daniel Whitenack?

  Ah.

• 8:45Speaker 03

  So in our normal venue, and we have a rotation of a few venues, we can hold about 120, 150 people in the room.

• 8:55Daniel Whitenack?

  That's awesome.

• 8:56Speaker 03

  Other venues, when we know we're gonna, we have gonna sell out, we have other venues that can hold 200 and even 300, depending on who's speaking.

• 9:05Daniel Whitenack?

  We need to, uh, we need to go to New York, Chris. Yeah, we do.

• 9:08Speaker 03

  Yes, and it's so much fun being here in person, but we do try to make the meetup a global community. I don't mean meetups in general, but I mean specifically the New York meetup. So we have the Slack that Daniel mentioned that is open to anyone from all over the world. And we now live stream all of the meetups. So if someone can't be in New York, they can see it live wherever they happen to be.

• 9:33Daniel Whitenack?

  That's fantastic. Quick question, has that impacted attendance in any way or has it just enhanced it?

• 9:40Speaker 03

  Our numbers for attendance have been pretty steady at sellout size before and after the live stream. So it looks like it's just people who can't physically be here or who couldn't fit in the room.

• 9:50Daniel Whitenack?

  That, that's awesome. I know I've appreciated being on the Slack and like I say, I've felt very welcome there. So we'll post some of those, uh, those links in the, in the show notes. Cause if people, even if they're not in, um, even if they're not in New York, it sounds like there's a lot of opportunities to get involved.

• 10:07Speaker 03

  Absolutely. There's always a way somehow that we just want everyone to be happy doing data science together.

• 10:15Daniel Whitenack?

  That's great and I really appreciate that. I mean, that's at the heart of what this show is as well, making data science and AI and machine learning accessible to people and practical. So, really appreciate your work there.

• 10:29Speaker 03

  Thank you.

• 10:30Daniel Whitenack?

  Yeah, so we're gonna get back to our, kind of the "our" AI language to steal some puns from your community. But yeah, I kind of wanted to get into a little bit. I think you're a great person since you've been at the forefront of data science for quite a while. And I know that we've talked about different AI techniques and deep learning and other things. And I also saw while the New York R conference was going on, some people tweeting out about how you were talking about deep learning and how we can understand it as kind of extreme nonlinear modeling. So, we've had some comments in our Slack, the change log Slack, about wanting to get

• 11:14Daniel Whitenack?

  some perspective from one of the guests on the different landscape of AI and ML techniques and how deep learning fits into that. So, I think you would be a great person to give us that context. Would you mind doing that?

• 11:28Speaker 03

  Absolutely.

• 11:31Speaker 03

  One of the key things about learning is what I've seen from teaching both private clients and up at Columbia is that you have to disambiguate terminology. And people often just drown in the different terminology. In fact, the talk I gave at the most recent New York R conference was comparing ML to deep learning in R. And the first thing I started with was just the vocab, because the different people from different fields will call the same exact thing with different words, like an intercept versus bias, or coefficients versus weights.

• 12:09Daniel Whitenack?

  Yeah, that was that was actually super confusing to me when I was starting to get in data science. I was like, oh, if I just would have known this was an intercept, then that would have made things so much easier.

• 12:19Speaker 03

  Exactly, exactly. It's almost as if you have people at different universities doing the research, not reading each other's papers and not knowing that these other terms exist.

• 12:28Daniel Whitenack?

  Yeah, yeah, for sure. So help us help us disambiguate some of that. You know, how do you see the the AI ML landscape in general and kind of some of the major categories within that?

• 12:42Speaker 03

  So I think first of the AI buzz, a lot of half the time when someone says they have an AI, they mean they have an if else statement.

• 12:51Daniel Whitenack?

  Yeah, that's that's true.

• 12:53Speaker 03

  And then beyond that, the majority, I do still believe in it from what I've seen in practice, the majority after that is a logistic regression. And half of what you see is coming out here is auto magical is the words they use. It's it's basically stuff that we've been doing for it's tools we've been using for 100 years. And I've been attending a number of Capitol Hill events about AI. And I heard someone make a really good point. He said, everyone's sort of up in arms right now about AI, but no one was concerned when we called it logistic regression.

• 13:29Daniel Whitenack?

  So yeah, that's, that's definitely, I think everyone wants to add AI to the tag of their project now, right?

• 13:39Speaker 03

  Yes, yes. There's been a run on dot AI domains.

• 13:43Chris Benson?

  They sure have. You can't get anything anymore. They've only been out for the last year or so.

• 13:47Speaker 03

  Exactly. It's very difficult. I managed to get a few domains for myself. I'm very happy about that I'm going to start using. But of course, you have to jump on the bandwagon.

• 13:57Daniel Whitenack?

  Yeah, I contemplated writing an AI that would would, you know, generate both startup names and then search for the AI domain and just register a bunch of them. And then I would raise venture capital and and sell out and retire.

• 14:15Chris Benson?

  How's that going for you, Daniel?

• 14:17Daniel Whitenack?

  You know, I didn't make it too far. It was mostly a theoretical construct, so.

• 14:22Speaker 03

  Is piedpiper.ai still available?

• 14:26Chris Benson?

  I don't know. I hope so.

• 14:28Speaker 03

  But to the question you asked, where the different parts fall in, I see AI in a few in a few different segments. There's automation, which could just be scripting, writing scripts. We've been doing a lot of work with some companies automating their data processes where people were literally copying and pasting from one spreadsheet to the other and now doing it all through scripts. And that's seen as AI by some people. It's made their lives easier.

• 14:55Chris Benson?

  That's interesting that we're kind of seeing kind of a crossover of, you know, I would only think of that as automation, but I guess if it is something you didn't have before, some people may define it that way in a very loose way.

• 15:08Speaker 03

  Exactly. It's like automated car production lines. That is a form of artificial intelligence, how to put the cars together. It's not thinking for itself, but it's doing tasks that humans used to do.

• 15:20Chris Benson?

  So, so let me, let me kind of ask a follow up on that. And, you know, how would, how do you think of AI? How does deep learning fit into that context? And, you know, how would you, how would you differentiate them?

• 15:34Speaker 03

  So the next step up from this automation is learning from data to make decisions for you. So first we have our linear models, our generalized linear models, including binary regression, Poisson regression, whatever you may have, to learn and do things not just based on hard set rules, but to learn from context. Now, linear models were simplification to make the math easy because they were invented 100 years ago or so. Then you get into nonlinear models, which at first were similar to linear models, but had a nonlinear construction. But then they got into trees and tree based models, whether they're, then they got into trees

• 16:16Speaker 03

  and tree based models, whether they are random forests or boosted trees. And boosted trees, especially XGBoost, were the darling of the machine learning community for years.

• 16:28Chris Benson?

  And Kaggle.

• 16:29Speaker 03

  And Kaggle. Almost every Kaggle competition was won using XGBoost.

• 16:34Chris Benson?

  Yeah, with a very narrow margin between the other similar methods, right?

• 16:40Speaker 03

  Right. So random forest was the rage and then XGBoost came on top of random forest. And then it turns out even with XGBoost, you could do a boosted random forest.

• 16:50Chris Benson?

  Boost all the booths.

• 16:52Speaker 03

  Exactly.

• 16:55Chris Benson?

  Nice.

• 16:55Speaker 03

  So then, beyond XGBoost, we have deep learning. And this is essentially a neural network that just has many layers. And neural networks have been around since at least the 50s. And they were cool, then they weren't cool, then they were cool again, then not cool, now they're cool yet again now.

• 17:13Chris Benson?

  And, and when you say a layer in the neural network, could you just explain a little bit what, what you mean by layer?

• 17:21Speaker 03

  So when you have a neural network, you have all of your input variables. And you have weights for them or coefficients for them. And you multiply those weights by the coefficients. And then you take that and do a nonlinear transformation. That is a layer. That is a set of now almost new inputs, not inputs, but they're new variables, if you could say so, but they've been transformed of a two-step process. And that is a layer. And you can repeat these layers again and again until you finally get to a point where you get to your output. And that's where the power of these neural networks are coming from today, having many of these hidden layers. And these are just transformations of your input variables. And the next hidden layer is a transformation of the first layer.

• 18:05Speaker 03

  And you can keep transforming on and on.

• 18:07Chris Benson?

  So let me ask, one of the things that I've, I've seen you talk about is extreme nonlinear modeling in reference to deep learning. And I was kind of wondering if you would kind of take us into what the difference is between those, or if it's the same, and, and how you see deep learning?

• 18:25Speaker 03

  So with a linear model, the reason we made it linear is because it was a simplification so that way they could do the math. But nowadays, we have more powerful computers and most things in life don't follow a linear relationship. They follow a nonlinear relationship. Now, when I say nonlinear, that could have different meanings depending on the technicality. But you can imagine if you had a cloud of points with a X and Y axis, instead of fitting a straight line through those points, if you fit a step function. Maybe for the first segment, it's about a third of the way up the Y axis. The second segment, the straight line would go two-thirds up. And the last segment, it would be back down to the bottom of the Y axis.

• 19:07Speaker 03

  And that's a simple step function that is nonlinear. It doesn't fit a nice straight line, doesn't even fit a curvy linear line. It fits a step function. And that's somewhat the idea behind a tree somewhat. And the ability to capture these nonlinear relationships, regardless of the method, allows us to really model reality better. That's why trees are really great. They have high predictive power and why random forests and boosted trees. That's also why deep learning is powerful because it is nonlinear. It has a lot of nonlinearities.

• 19:48Speaker 03

  So when you're going from your inputs to your first hidden layer and then on to subsequent hidden layers, there are two steps. There is a matrix multiplication of the inputs by their weights or coefficients, and that's linear. If you just did that, a deep learning model would just be a linear model. You could even stack many more layers, and if you just did these multiplications by the weights, it would just be a series of linear models, which would become one large linear model. And then you essentially have a straight line or a curvy linear line. But it's that next step at each layer, the activation function.

• 20:32Speaker 03

  That is a nonlinear function you are applying. So whether it is a tanh or it's a ReLU or a sigmoid, which is just a fancy word for inverse logit, regardless which one you're doing, you are doing a nonlinear transformation. And that puts a nonlinearity in your model, which allows you to capture more complex relationships. If you do more layers, you have more nonlinearities, so you can capture really interesting separations between your data.

• 20:59Daniel Whitenack?

  So yeah, that's a really, really great context. And I wonder, so like, sometimes when I'm thinking about these problems solved by deep learning, I think about them in terms of, you know, I know that there must be these relationships between what I'm putting in and what I want to get out, but I have a really hard time understanding what those relationships are. In your thought process, when you're doing this sort of deep learning technique, does that put you kind of further away from actually getting insight into those relationships? Or does it, I mean, does it really matter at that point?

• 21:36Speaker 03

  That's an interesting question. Does it matter? Depends on your goal. If you're doing a study on a medical trial, you really want to know what's happening because you need to know is the drug helping or hurting. You could do that with prediction, but you really want to know what's happening with the treatment. What type of effect is it having? Is it a large effect? Is it a small effect? And while there are ways to get that from purely predictive methods like trees and deep learning, it's not as explanatory. If on the other hand, you're just trying to make a prediction and you don't necessarily care why, then it doesn't matter. But the idea of understanding it, though, even with a logistic regression that is complex

• 22:19Speaker 03

  with lots of interaction terms could be hard to interpret. So yes, we are losing more of that interpretation with a deep learning model. But depending what you're doing, that's not terrible if you don't need to understand what's happening.

• 22:35Chris Benson?

  Gotcha. So as we've been kind of talking about, you know, how we're approaching AI and deep learning specifically and such. One of the things that comes to mind is I'm curious if you're actually using deep learning with your clients. And if so, if you can kind of tell us what that looks like a little bit.

• 22:53Speaker 03

  Yes, we have a few clients we do deep learning with. In fact, we became Nvidia partners because they are really pushing the forefront of deep learning. They have a vested interest with their GPUs. So we work with some of our own clients and some of Nvidia's clients on specifically deep learning problems.

• 23:11Chris Benson?

  So let me ask another question, especially coming as a newbie to the R community. What are some of the strengths that R have that you specifically find really help you in the development of AI or ML technologies?

• 23:28Speaker 03

  R was written from the ground up as a data language. It was meant for handling data of different types, whether they are numbers or text or dates or logicals. It was designed for data. And beyond that, it was designed for statistics, which is data science, ML, AI by another name. It was meant for doing matrix algebra. It was meant for mathematical programming. So anything involving data just comes so naturally to the R language that it's a joy to work with.

• 24:00Daniel Whitenack?

  So with respect to that, I mean, I think my kind of stereotypes when I, before I was involved with the R community was that, you know, R was used by these kind of people at universities and they wrote kind of one-off things to do nifty visualization stuff and data munching, but it wasn't really like a production language, you know, that people are using for, for AI in production at companies. But that's not quite true, is it?

• 24:29Speaker 03

  Not true at all. There's a number of companies that I know are using in production, which I can't name here. They're my clients. But I've also seen other companies, not my clients, who do use R in production on a daily basis. And it is a full, robust language, just like any other.

• 24:47Daniel Whitenack?

  Where do you see the largest kind of interest in R industry-wise, you know, whether that's like finance or healthcare? I know you're in New York, so maybe like finance is a big one. Do you also kind of like, what do you see as the main industries that have that interest? And kind of, is it also geographically distributed? I know you're in R and maybe there's difference kind of on the West Coast or something.

• 25:14Speaker 03

  So for industries, I see the most adoption from personal experiences in finance, in pharmaceuticals and healthcare, and in defense.

• 25:24Daniel Whitenack?

  Interesting. I didn't expect defense. Is that just because you think that's because those are communities that are being fed from kind of R-heavy academic fields? Or what do you, what do you think is the reasoning behind that?

• 25:40Speaker 03

  It really depends how you come up. If you come up thinking about math and statistics, you're going to be an R person. If you come up thinking about engineering, computer science, you're going to be a Python person. So a lot of the government agencies I've seen, whether they are defense-related or even stuff such as the National Institutes for Standards and Technology or the Census Bureau, a lot of them are going to be trained as statisticians, mathematicians. So R comes naturally to that community.

• 26:08Daniel Whitenack?

  Yeah, that's, that's fine. So you mentioned NIST and I just had like a total flashback when I was in grad school, I think. So I had a part-time appointment there and, and I think all I did was like spill nano tubes all over the floor everywhere at NIST. I was, that's where I found out that I should not be in a actual lab, that I should just work on a computer.

• 26:33Speaker 03

  Yeah, no wet lab work for you.

• 26:34Daniel Whitenack?

  Yeah.

• 26:35Chris Benson?

  So I have a follow-up question to that last answer you gave us. You kind of talked about how depending on what field people are in, they naturally gravitate to, to this language or that. And, you know, but we're kind of in this age of polyglot programmers. And so if you're-

• 26:55Daniel Whitenack?

  That is a hard word to say.

• 26:56Chris Benson?

  It's a hard word to say. So if you're one of those people and you're, you start in, maybe you start in R and then you also pick up Python or, or anything, could be Julia, could be Go, could be whatever. Do you have any advice on, on if you have multiple languages in your capability, when you might go to R and when you might not go to R and go to a different language instead? Any thoughts on that?

• 27:23Speaker 03

  Yeah, I see going to R when you want to do, get something done quickly and it's really data machine learning focused. It's, like I said, it's natural to be used for that stuff. If you need something that's blazing fast, then you're going to do C++ or Go. If you need something that is out of your wheelhouse in R that is more, let's say building a web server, even though you could do that in R, you might go to do that in a language that's more natural for a web server. So in my mind, it's really, what is your goal? What is your broader ecosystem? Depends how does your company work? Do you have a pipelining tool that can pull from different resources or does everything have

• 28:07Speaker 03

  to be monolithic? So it really depends on all those factors around you and what the task at hand really is.

• 28:15Daniel Whitenack?

  Yeah, and I think even to that point, there's a lot of interaction with other communities and programming languages as well. I know that you're you're pretty good friends with Wes McKinney, right?

• 28:26Speaker 03

  Yes, yes, we are. We are very good friends.

• 28:28Daniel Whitenack?

  Yeah, and I remember, I think it was at the R conference there in New York, that that he was giving a talk about some of these these layers that actually like bridge the gap between the data science and AI languages, like Apache Arrow and other things. Could you could you kind of share some of those those kind of intersection points? I think those are really exciting.

• 28:53Speaker 03

  Yeah, so Wes has been working on Apache Arrow, which as many things, but one of the one of the key points to it is an interoperable data frame. So you have a data frame in R, you save it to disk, and you could open that binary file in Python, or Julia, or any other language that supports it. And that's fantastic. Even deeper than that, all these interpreted languages, particularly R and Python, are really just high level bindings for C and Fortran.

• 29:22Speaker 03

  So the same libraries are doing all the work, and there's just a different flavored wrapper on top of it to suit different people's needs.

• 29:29Daniel Whitenack?

  I think you just shook some people's worlds that maybe didn't know that they were using using Fortran.

• 29:35Speaker 03

  And when Daniel said he did Fortran, I always say that anyone who knows Fortran is awesome.

• 29:42Daniel Whitenack?

  Well, I don't know that I would say I know Fortran. When I quote unquote, wrote Fortran, I mostly wrote Python around Fortran, because I didn't want to actually touch the Fortran for fear of breaking all sorts of large code bases that that were beyond my scope. But occasional dabbling, maybe. I'll leave it at that.

• 30:03Chris Benson?

  Now, Jared, I got to say, you just made my day, though, because my mother, who is now retired, taught Fortran at Georgia Tech back in the 1980s. And the fact that she is that compliment you just offered for Fortran, she's gonna love that. I might even get her to listen to the podcast now.

• 30:18Speaker 03

  Oh, that's fantastic. I love hearing that.

• 30:20Daniel Whitenack?

  I I'm really happy that our that our guest unknowingly complimented Chris's mom.

• 30:28Chris Benson?

  Can't beat that.

• 30:29Speaker 03

  No, that's the best.

• 30:32Daniel Whitenack?

  Yeah, that's awesome. Yeah, that's great to hear about those kind of intersection points. I think more and more I'm seeing, which I'm really happy to see, I'm seeing kind of, you know, hopefully less of these language war sort of competitions or discussions on Twitter and other places and more intersection points between language communities, whether that be Apache Arrow, like you were just saying, or whether that be something like the ONIX neural network formats that exchange between different frameworks and all of that is really exciting to me.

• 31:10Speaker 03

  As long as you're doing this machine learning, doing AI, doing data science, people shouldn't care how you're doing it. Do whichever way works best for you and for the problem you have and use the tools that feel comfortable to you.

• 31:23Chris Benson?

  So I have a question that's especially for me and maybe a few listeners out there in the same boat as me. Since I'm brand new to R and haven't used it in the past like you guys have, I want to ask you, you know, for AI ML context here, how should I on day one kind of get started with machine learning in mind? What do you recommend?

• 31:48Speaker 03

  Well, of course, the first step if you're just learning R, you have to get a copy of my book, R for Everyone, obviously.

• 31:53Chris Benson?

  Perfect.

• 31:54Speaker 03

  Ching! Exactly. You have to do it. I mean, there's no other way.

• 31:58Daniel Whitenack?

  Done deal. Well, we'll for sure put that link in the show notes. I even have evidence. I've seen the lines at conferences to get Jared's signature on his book and it's, it must be awesome if the lines are that long. So.

• 32:15Speaker 03

  It's always very flattering when I see a long line of people lined up to come get my autograph. I'm like, I could do this all day long. Then beyond that, you want to start getting more specifically into ML. Of course, you have to show up to the meetups and the conferences. That's your in-person experience. And that's not just for the knowledge, which you do get a lot of knowledge, but it's to be around other like-minded people.

• 32:36Daniel Whitenack?

  And for the pizza.

• 32:38Speaker 03

  And for the pizza. Yes, we do pride ourselves. In fact, every month, we try to get pizza from a different place and we have a data set running. People rate the pizza. You can go pull that data. It's a live feed and see how it's going over time.

• 32:51Daniel Whitenack?

  I'm assuming that you haven't brought in some Chicago-style pizza yet, which is my favorite, but probably less exciting in New York.

• 33:00Speaker 03

  So you mean casserole?

• 33:05Daniel Whitenack?

  I think we should move on to the next question.

• 33:07Speaker 03

  Yes, but finish answering the question about statistics. Then, once you've already got a basis in R, the question becomes, do you already know the math, the statistics? If you don't, there's great books out there written by Andrew Gelman and Hasty Tip Johnny and Friedman and Jennifer Hill. Those are great books to learn about the statistics. And then, if you have a grounding in that, it's about how do you do it in R? Now, how do you define ML? Well, you want to learn linear regression. That's a simple one-line of code. Then if you get into penalized regression and XGBoost, you could do that all natively in R, or you could use a package called Keras. And Keras was a unified interface for machine learning in R

• 33:51Speaker 03

  written about 10 years ago. And it's just gotten better and better since then. And that was written by a guy named Max Kuhn, and he has a companion book called Applied Predictive Modeling. And then, when you're ready to get into the deep learning part of your R experience, there's the older packages like neural net, and there's newer packages like MXNet. And J.J. Leir, the head of our studio, wrote a couple of packages called TensorFlow and Keras. And he wrote a book to go along with that called Deep Learning in R. So if you go through all of these books and go through using the functions yourself to Keras, you'll have the whole spectrum of doing AI all within R.

• 34:32Daniel Whitenack?

  Yeah, and I think, you know, I had a great time learning some R. I'm by no means an expert, but I think, like Jared kind of already mentioned, it's a pretty quick way to get from nothing to something working in a short period of time. And I think that makes it a lot of fun. But you mentioned TensorFlow, Jared. What is the state of interaction between R and TensorFlow? I know it's kind of been a long time coming. For example, in the Go community where you can do inference. You know, it's recommended that you do inference in Go with TensorFlow, but not necessarily training. What's kind of the state of interaction between R and TensorFlow? And are there other kind of?

• 35:16Daniel Whitenack?

  Larger frameworks that are integratable with R, like, you know, maybe PyTorch or other things?

• 35:25Speaker 03

  So with the TensorFlow and Keras packages, you get the full functionality of TensorFlow and Keras.

• 35:31Daniel Whitenack?

  That's awesome.

• 35:32Speaker 03

  Yeah, it's really amazing. I sit on my computer, even on my, actually, my Windows laptop, and I can build TensorFlow models thanks to the Keras interface in R.

• 35:42Daniel Whitenack?

  That's awesome. Who did that work?

• 35:44Speaker 03

  So it was mainly J.J. Lehrer, he's the head of our studio, and Francois Chollet was also involved, and the team at Google.

• 35:55Daniel Whitenack?

  That's awesome.

• 35:56Speaker 03

  It's really amazing. Now, it has TensorFlow as the default, but you can plug in other compatible Keras frameworks. So it really gives you a broad spectrum of what you can do, though I would say that 99% of the people using Keras in R are using TensorFlow as a framework underneath.

• 36:14Daniel Whitenack?

  That makes sense.

• 36:15Chris Benson?

  That's pretty cool. So I guess as you look forward in R at this point, and as we are surging forward for the years to come in AI and ML and stuff, what are you excited about right now in the R community, and in your own projects, for that matter, where R is intersecting AI and ML?

• 36:33Speaker 03

  So it's kind of funny. You see other communities, and they're super excited about all the different machine learning stuff they can do. In the R community, it's sort of larger, like, been there, done that. R's been doing it for decades.

• 36:45Daniel Whitenack?

  They're already on to better things.

• 36:48Speaker 03

  Not to say better things, but different fun things. Like, we're all super excited about using R Markdown to automate slideshows and reports, or HTML widgets to have interactive JavaScript embedded in your R report, or different types of graphics, or data manipulation, or network analysis, all this other fun stuff that, like, yeah, we've been doing machine learning forever. Look at all this other cool, fun stuff.

• 37:09Daniel Whitenack?

  Yeah, and I guess at this point, I mean, you have awesome support for things like TensorFlow, like you already said, and you know, those sorts of interactions with JavaScript widgets or whatever it might be are really, I think they're really interesting.

• 37:23Speaker 03

  And they're really, really cool and fun stuff because let's be honest, the vast majority of time a data scientist spends is not doing all the really cool modeling that we all want to do. It's doing the data prep, the manipulation, reporting, graphing, and that's 80 to 90% of the job now. Because now it's become so easy to do the modeling, to do the true AI part, that everything else takes up so much time.

• 37:49Daniel Whitenack?

  So I have a, I have a confession that I need to make probably early on in this podcast, and that's that all of this data munging and cleaning and all of that, I really, really enjoy that.

• 38:04Speaker 03

  There's something fun about it. It's like this problem solving, getting to patch things together.

• 38:08Daniel Whitenack?

  I don't know what makes me weird in that way, but I just, I could spend just days heads down cleaning data. I really like it.

• 38:18Speaker 03

  We need more people like that.

• 38:20Chris Benson?

  I'll tell you what, I know. I don't know whether to praise you, Daniel, or just to say, wow, that's, that's a little bit crazy.

• 38:26Daniel Whitenack?

  Well, yeah, pair me with, with one of you guys and I'll, I'll do the data cleaning and you can, you know, add an awesome JavaScript widget powered by TensorFlow or something.

• 38:39Speaker 03

  It's funny you make that joke, but you know, there's now JavaScript bindings for TensorFlow.

• 38:44Daniel Whitenack?

  Yeah, that's, that's a really, a really interesting topic, and I think that there's like some very subtle but really important implications of things like that. And that's, you know, around like privacy and other stuff like that. If you're, if you're actually embedding a model in JavaScript and running it on someone's, you know, device in their browser or whatever it might be, and maybe even updating a model in the browser, you know, then data never has to leave that person's device. I think that's like a really important and interesting implication of stuff like that.

• 39:17Chris Benson?

  I sense a show coming up on that topic.

• 39:20Daniel Whitenack?

  I would love to have that. I saw the, so there's a demo, maybe we'll, we'll find the link and put it in the show notes of the recent TensorFlow Dev Summit. They had a web app that you could play Pac-Man with like your head movements running in the, in the browser and you would just have to calibrate it by moving your head, you know, a certain number of times. And it would actually, you know, online update the model and then use your particular head movements to control the game.

• 39:49Chris Benson?

  And then after you played the game, they had chiropractors and, and people giving massages, you know, just to.

• 39:54Daniel Whitenack?

  It's probably not a game that you're going to want to, you know, want to play all night, but. Jared, so that's a, that's a lot of interesting stuff, but before we go too far down the Pac-Man hole, is there, is there anything else that, that you wanted to mention that, that you're excited about or things coming up?

• 40:13Speaker 03

  So something I'm seeing a lot in the data science community in general, but particularly the R community, is what to do with these models after you've built them. And in the past, I came up with all sorts of workarounds. I'd take a model, save it as a binary and then have an R session running and people would have to interact with that using R script. And now it's becoming so easy through various different tools for other people to now consume the results of the model, particularly with predictions or scoring, depending on the word you want to use, or even inference as the deep learning community uses, even though that means something else in the stats community. It's now become so easy with different tools such as the Plumber API package in R, which turns your R scripts into APIs,

• 40:57Speaker 03

  or even third-party solutions. There's a company out there called Algorithmia and they take your models regardless of the language. Then you build a API that anyone can hit. So these tools, we can productionalize our model so easily and I could write my code in R. I could use either Algorithmia or Plumber to create a simple RESTful API, have it running on a server or even a microservice. Then someone else can go hit that and get the results of my model. So really. Creates this accessibility and it's so exciting seeing this happening faster and faster these days.

• 41:31Daniel Whitenack?

  Yeah, so I was actually, it's funny that you mentioned that because this morning, on my, on my drive in, I was listening to the change log episode, which is our, our overlords podcast, which you should definitely check out. But you know, our, our kind overlords podcast and they were talking to, I think it was the VP of AI at Microsoft. We'll link the show in the show notes, but he was using this terminology around democratizing AI. And essentially that what we're seeing built now with these tools, like you mentioned, like, like plumber and algorithmia. And I think other things like machine box, like we had on another episode, is actually creating a layer,

• 42:15Daniel Whitenack?

  a new layer in the software stack, which is making machine learning and AI techniques, you know, available to, you know, developers that might not have any interest or ability to go deep into the math of the model that they're using.

• 42:31Speaker 03

  Yeah, it's, it's making our results accessible to other people to consume. And that makes it easier to, hey, they can do their job and I can do my job. And then they, we both get a benefit from each other's work.

• 42:44Daniel Whitenack?

  Yeah, exactly. And you can say, hey, we're exposing the ability for your program to recognize things or understand speech or whatever it is. And all they have to do is think about implementing that functionality within their software stack now via a nice API, like a, like a REST API or whatever it is.

• 43:04Speaker 03

  Exactly. And they could be that much more efficient and consume these things that they had no chance of before. They would have had to spend all the time building it. It's specialization in a very great way.

• 43:16Chris Benson?

  Well, I am super psyched up, Jared. I am gonna have to, as soon as we finish up recording, I'm gonna have to go grab your book and, and start my own journey into R and try to catch up, at least with Daniel at some point here. And so, thank you so much for, for taking us through this and for me kind of giving me a path forward on how I can start learning R and using it for AI and machine learning.

• 43:40Speaker 03

  Great. Thank you for having me. It's been real fun for me, too.

• 43:44Daniel Whitenack?

  Awesome. Thanks.

• 43:47Speaker 00

  All right. Thank you for tuning into this episode of Practical AI. If you enjoyed the show, do us a favor, go on iTunes, give us a rating, go in your podcast app and favorite it. If you are on Twitter or social network, share a link with a friend, whatever you gotta do, share the show with a friend if you enjoyed it. And bandwidth for changelog is provided by Fastly. Learn more at fastly.com. And we catch our errors before our users do hear changelog because of Rollbar. Check them out at rollbar.com/changelog. And we're hosted on Linode Cloud Servers. Head to linode.com/changelog, check them out, support this show. This episode is hosted by Daniel Whitenack and Chris Benson. Editing is done by Tim Smith. The music is by Breakmaster Cylinder. And you can find more shows just like this at changelog.com. When you go there, pop in your email address, get our weekly email

• 44:32Speaker 00

  keeping you up to date with the news and podcasts for developers in your inbox every single week. Thanks for tuning in. We'll see you next week.