The influence of open source on AI development

The All Things Open conference is happening soon, and we snagged one of their speakers to discuss open source and AI. Samuel Taylor talks about the essential role that open source is playing in AI development and research, and he gives us some tips on choosing AI-related side projects. Sponsors: DigitalOcean https://do.co/changelog – DigitalOcean now offers three managed databases https://www.digitalocean.com/products/managed-databases/ — PostgreSQL, MySQL, and Redis. Get started for free with a $50 credit. Learn more at do.co/changelog https://do.co/changelog. The Brave Browser https://brave.com/cha913 – Browse the web up to 8x faster than Chrome and Safari, block ads and trackers by default, and reward your favorite creators with the built-in Basic Attention Token https://basicattentiontoken.org. Download Brave https://brave.com/cha913 for free and give tipping a try right here on changelog.com https://changelog.com. All Things Open https://allthingsopen.org/?utm_source=changelog – Exploring open source, open tech, and the open web in the enterprise. Raleigh, NC — October 13-15, 2019 Featuring: Samuel Taylor – Website https://www.samueltaylor.org, GitHub https://github.com/ssaamm, LinkedIn https://www.linkedin.com/in/samuelgt, X https://x.com/samueldatatChris 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: Tweet and mention @Changelog https://twitter.com/changelog or @PracticalAIFM https://twitter.com/practicalaifm for a chance to win 1 of 5 free passes to the conference! “I want a free pass to @AllThingsOpen because…”All Things Open https://allthingsopen.org/Samuel’s All Things Open talk description https://allthingsopen.org/talk/using-open-source-tools-for-machine-learning/Episode #17 - Fighting bias in hiring https://changelog.com/practicalai/17Jupyter https://jupyter.org/Pandas https://pandas.pydata.org/Scikit-learn https://scikit-learn.org/stable/Upcoming 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

  Being with your 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 Leno cloud servers. Head to leno.com slash change log.

• 0:16Speaker 00

  This episode is brought to you by Digital Ocean, the simplest cloud platform out there. And we're excited to share they now offer dedicated virtual droplets. And unlike standard droplets, which use shared virtual CPU threads, their two performance plans, general purpose and CPU optimized. They have dedicated virtual CPU threads. This translates to higher performance and increased consistency during CPU intensive processes. So if you have build boxes, C I C D, video encoding, machine learning, ad serving, game servers, databases, batch processing, data mining, application servers, or active front end web servers that need to be full duty CPU all day, every day. Then check out Digital Ocean's dedicated virtual CPU droplets. Pricing is very competitive, starting at 40 bucks a month.

• 1:01Speaker 00

  Learn more and get started for free with a $100 credit at do.co slash change log. Again, do.co slash change log.

• 1:21Speaker 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 slay 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 into the show.

• 1:46Daniel Whitenack?

  Well, welcome to Practical AI. This is Daniel Whitenack. I'm a data scientist with S I L International. And I'm joined by my co-host, Chris Benson, a Principal AI Strategist at Lockheed Martin. How are you doing, Chris? I hear you just got off of a plane.

• 2:03Chris Benson?

  Yeah, I just arrived in in London via Heathrow and just got into the hotel room in time to record here. So looking forward to it. I'm in London because tomorrow I will be on a panel representing Lockheed Martin at the Royal Academy of Engineering for a panel discussion on artificial intelligence of all things.

• 2:23Daniel Whitenack?

  Oh, wow. That sounds intense. I don't know if I've ever been involved in any sort of royal.

• 2:29Chris Benson?

  It's scary. The word, the word, put royal in front of anything and it's either big or scary or both.

• 2:35Daniel Whitenack?

  Right, exactly. I just, uh, yeah, I don't know if there'll be anybody in like robes or wigs or anything in the audience. So that's kind of what I have pictured.

• 2:44Chris Benson?

  I'm hoping they'll let me do that, you know, we can do, you know, kind of like the old movies, you know, we'll all be up on the panel talking about AI with our wigs on and stuff. That'd be perfect.

• 2:53Daniel Whitenack?

  Yeah. Well, on a slightly different note, there's another event coming up relatively soon, uh, which is a conference called All Things Open, which I've been to, uh, once I spoke there. It was a great experience. Um, a conference kind of focused on a lot of different things other than machine learning and AI, but kind of centered around the open source world. And one of the people talking there this year is Samuel Taylor, who is a data scientist at Indeed. And he's talking about using open source tools for machine learning. And so we definitely thought that that was practical for those out of us trying to do practical AI. So we've got Samuel Taylor with us this week to, uh, to talk about it. Welcome, Samuel.

• 3:37Speaker 03

  Thank you all so much. I'm, uh, really excited to be here, really interested to be part of this project that y'all have to make AI more practical and more, uh, easy to use for, for people.

• 3:48Daniel Whitenack?

  Thanks. Yeah, it definitely seemed like your, your talk was in that vein and, um, we'll, we'll get to the subject of the talk, but maybe to start out, could you just give us a little bit of information about your, your background and how you got into data science and machine learning?

• 4:03Speaker 03

  Of course. So from some accounts, I've just always liked computers. There, there's a picture of me when I am, I think, seven or eight years old, sitting at the family computer and I have my hand on the mouse, playing some point and click adventure game. And I guess I just never grew out of that. And just always liked computers and had really great parents who encouraged me to pursue. Math and programming. And I was able to learn a lot of that and practice a lot of that leading up into high school. And at some point in high school, I saw a documentary on PBS. I think it was like a Nova documentary. And they were talking about machine learning, I guess, because they had an example where they showed a computer a bunch of images of English letters.

• 4:45Speaker 03

  And then it could tell with a new picture if it was an A or a B or a C or whatever. And that just blew my mind.

• 4:51Daniel Whitenack?

  You gotta love those Nova documentaries.

• 4:53Speaker 03

  Oh, yeah. Yeah, totally, totally. And it just, like I said, it just blew my mind. And then when I got into college and started studying computer science and realized like, I can learn how to do that. That's, that's really cool. I took some classes in that. And then after I graduated, started doing more like software engineering stuff and data engineering. And then at my current company, Indeed, we have this internal transfer program where I was able to transfer to a data scientist role at the company, which has been really good. So I guess that's how I ended up in data science.

• 5:22Daniel Whitenack?

  Congratulations. Yeah.

• 5:24Speaker 03

  Yeah, thanks.

• 5:25Daniel Whitenack?

  Do you feel like, I'm kind of, I'm always intrigued by people's journeys and how it influences how they think about data science problems. Do you feel like your sort of initial emphasis on software engineering and building up that side of your, your expertise, how do you think that's influenced, you know, your work in, in data science?

• 5:44Speaker 03

  That's a wonderful question. I think it is a, it plays a huge part in the way that I think about these systems because I recognize that I am never going to be as good at applied statistics as someone with an astrophysics PhD, right? And I just know that that's not, not my strength. And so what I try to do is then understand these algorithms like I would any other algorithm and try to treat them like any kind of other engineering system and treat them with the same kind of rigor that I would in that way. The other way it's been helpful is that it kind of sets me up well to try to, try to bridge the gap between some of these people who are brilliant statisticians and really understand data deeply. And then, you know, sometimes, especially if they're coming straight out of school, might not have as much experience in the, the software engineering side of things. And so.

• 6:29Speaker 03

  It can create this really useful kind of knowledge share where I'm able to help some people with kind of scaling up at software engineering and they're able to help me get better at the stats side of things and introduce me to stuff I'm not familiar with. And so I think it ends up just being really good to have that diversity of backgrounds. It's very, I think it's been really good for me.

• 6:47Chris Benson?

  So now that you've told us a little bit about your background and the fact that you're now at Indeed, to kind of give context to our conversation, can you tell us a bit more about Indeed? I know, obviously, Daniel and I are very familiar with it, but there might be some people out there who haven't used it and let them know what it is.

• 7:01Speaker 03

  Absolutely. Indeed, we are the world's number one job site and the main thing that we emphasize, I think I've seen this on at least half of the presentations that people give even internally, is our mission is to help people get jobs and that's what we try to do every day. We have like orange chairs in our conference rooms to remind us that the job seeker is always the important thing and we need to try to do things to help them. So at its core, Indeed has sort of a search product that is well-liked where people will go and you can search like data scientist jobs in Austin, Texas and it'll have a list of jobs and there's all sorts of filtering and stuff that you can apply to find the right job for you.

• 7:41Chris Benson?

  So also wondering here, kind of how long has Indeed been investing in data science and machine learning and what's their primary focus in that area?

• 7:51Speaker 03

  Definitely. So as with any other large internet company, there's a mound of data that you get just running an internet company and it's obviously if you can leverage that well, then you can do a lot of great stuff to help people get jobs better. And I wasn't involved at the creation of the team, so I can't really speak too much to the early parts, the early days, but it's been at least several years of investing in data science at our company. Some relevant use cases, for instance, salary estimation is one that comes up fairly often and it's nice to be able to, for job seekers, it's nice to be able to have some expectation of what the salary will be for a given job. Another example that ends up being useful in a variety of ways is

• 8:36Speaker 03

  Training models that can determine how good of a fit a certain job is for a certain job seeker, and that can be leveraged in a few different ways. But with the data that we have, we're able to come up with these useful models that we can apply in several ways, which has been really nice. And I think, I mean, by empirically testing, we've found has really helped us be able to help people get jobs better.

• 8:59Daniel Whitenack?

  So at Indeed, is it kind of, is data science scattered throughout the teams that are working at Indeed, or is there kind of one data science group that consults on different projects? I know that those are two kind of common, common patterns that I've seen that probably each have their advantages or disadvantages. But at somewhere like Indeed that I'm sure has a lot of different projects going on, how does that work?

• 9:25Speaker 03

  Of course. So at Indeed, we try to practice something that we call full stack data science, where one individual is in charge of everything from coming up with an idea of some model that we could build through to gathering the data up about it and generating labels for it in some useful way, training the model, doing all the hyperparameter tuning, and then finally, you know, getting it deployed to production, writing that production code, monitoring it after the fact, testing it. So that's sort of the model that we try to practice, and we find a lot of benefit in that. And that enables us to have, you know, a group of data scientists who all end up placed in various teams and are able to really provide a lot of value to an individual team in that way, because they have a wide

• 10:08Speaker 03

  variety of skills and are able to get something all the way through from idea to actually in production.

• 10:14Daniel Whitenack?

  Full stack data science, I've definitely, I've heard that term a few times recently. I don't know if it was from Indeed, but I think that that's, that's starting to be used a little bit more widely. Have you heard that, Chris?

• 10:26Chris Benson?

  I have. It's, it's becoming a popular buzzword now. So we have formally introduced it on the show here. So we are, we are in place.

• 10:35Daniel Whitenack?

  I don't know if I can be considered of, I think, maybe based on your, your description, Samuel, I, I hope that I'm considered somewhat full stack, but I, I feel a little bit like cringe calling myself a full stack data scientist. I feel much better if someone else called me that. But anyway, not to get sidetracked.

• 10:54Chris Benson?

  Yeah. Well, actually, before you completely eliminate the sidetrack, it's kind of funny that you say that because, you know, we came from software development. It sounds like all three of us have come from there. And, you know, once upon a time, I did think of myself as a full stack software developer. So I, I wonder if we're going to grow into the, the sense of being full stack, you know, machine learning engineers, data scientists, whatever.

• 11:14Daniel Whitenack?

  Do I have to learn more JavaScript to be a full stack data scientist?

• 11:19Chris Benson?

  You have to do React with, with your AI from, at that point, you know?

• 11:23Daniel Whitenack?

  I'll, I'll look into that.

• 11:24Chris Benson?

  There we go. So, hey, you know, I'm trying to recall back, we had, um, from HireVue, we had Lindsay, I'm trying to remember, I'm going to butcher her name.

• 11:32Daniel Whitenack?

  Zulaga.

• 11:33Chris Benson?

  Zulaga, that was it. Uh huh. And, um, I think that was going back in our late teens, maybe up to episode 17, if I recall. She was talking about bias in hiring and data. And so, you know, Samuel, I'm kind of wondering, is that something that Indeed is working on as well? And, you know, it's a pretty big issue out there. And if you are working on it, kind of where is Indeed taking it as a company?

• 11:56Speaker 03

  Of course, that's a, a huge issue. Like it's to the point where that kind of stuff is coming up at conversations at the national level, you know, in, in the presidential debates, people are caring about bias in algorithms and bias in data. And there are people at Indeed working on that. I really can't speak to what they're doing. I just am not super familiar with what they're doing. One area that I have found interesting that can be somewhat related here is working in cases where you have really imbalanced data sets, where you are drawing from, where certain parts of the population that you're looking at might just be really underrepresented in your data set, and trying to come up with useful techniques for correcting for that, or for making sure that your model is still doing well on those subsets of the data that are underrepresented. I find really interesting

• 12:39Speaker 03

  because I feel like that comes up all the time. You'll have a case where For instance, like your target variable could be 1% or .0001% of the smallest class, and then everything else is the majority class. And I found that to be a really interesting problem to try to attack. I don't think it's directly related to the issue of bias in machine learning, but I think there could be some benefit there to be had, for sure.

• 13:02Chris Benson?

  You know, I think that is probably the epitome of, you know, the type of area in terms of bias and even its extension into, you know, kind of the newfound field of AI ethics. And that is the one thing all of us are grappling with. I think no matter where we're at is the fact that, you know, we have these data sets and we're trying to create great models. And that's just a universal challenge. Just about everyone I ever talked to says that.

• 13:25Daniel Whitenack?

  Yeah, definitely. It's hard to grapple with, for sure. Yeah, every time I encounter that, it makes you stop and kind of take a step back and really think through your process and really how your data was generated, what the implications are of different sampling techniques and all of those things. Yeah, it's hard every time I encounter it, it seems like. So Samuel, given that you work at Indeed and given that like data scientists are in demand and, you know, machine learning, AI is all the rage, do you have any sort of general, maybe just from your personal experience or patterns that you've seen, do you have any recommendations around, hey, I'm looking for an AI job or I'm looking for a data science

• 14:09Daniel Whitenack?

  job, what are maybe some good things to avoid or some good things to do generally as you're kind of going through that hiring process?

• 14:17Speaker 03

  Definitely. So obviously, I can only really speak to what I know and what I know is how I got into this. And I think one of the biggest things that helped me get into this field was being able to work on sort of side projects, you know, either after work or on the weekends or something. And I know that's not always an option for people who are busy or have kids. But if you can, if you do have the chance to do that, I think that's a really strong way to both just develop the skills and run into problems that you're going to run into in real life, but also to have a sort of portfolio to show to people and say, hey, look at all the cool things I can do with this. People are a lot more likely to take you seriously. If you have some sort of example that you can show them of, "Here's the thing I built. Look at how neat this is."

• 15:01Speaker 03

  And it can be a really good way to get on someone's radar if you can send them a link, for instance, to a website that you made that does some cool machine learning-y thing, and they can play around with it and be like, "Oh, this is fun." And even if that website isn't super complicated, they still might end up being, you know, it helps set you apart from the rest of the crowd.

• 15:33Speaker 04

  This episode is brought to you by Brave. The Brave team is on a mission to fix the web by building an open-source, privacy-focused, and performance-oriented browser, browse the web up to eight times faster than Chrome and Safari, block ads and trackers by default, and reward your favorite creators with the built-in basic attention token. Yes, you heard that right, a real-world use case for blockchain. Download Brave for free using the link in the show notes and give tipping a try on changelog.com.

• 16:14Chris Benson?

  So Samuel, let's turn toward All Things Open. I'm familiar with it, and I know Daniel is, but there may very well be people in the audience that aren't. Could you kind of tell people what All Things Open is about? And, you know, what organization backs it, that kind of stuff?

• 16:30Speaker 03

  Yeah, definitely. So All Things Open is this just massive conference that takes place in North Carolina, in Raleigh, North Carolina, and has several thousand, I think it's in the three or four thousands of people that show up to this thing. So I haven't gotten to go before. I'm really excited to because conferences are always just a blast. Anytime you get around like 3,000 other nerds and you're all just like there to celebrate the nerdy things that you like, it's always a good time. There's always really interesting people, so I'm really excited to get to go.

• 17:05Daniel Whitenack?

  Yeah, and I think especially, I don't know if this has been your experience, Samuel, but like at conferences that are very open source focused and sort of have a community vibe, there's just a lot of excitement there and, you know, always interesting people to talk to and really, really interesting kind of random but awesome projects that probably wouldn't get highlighted at like very, very expensive industry conferences. Is that your impression as well?

• 17:33Speaker 03

  Oh yeah, definitely. I have had, I mean, I've gone to some other more community, open source-y kind of conferences and the people you run into, you'll just sit down with someone and not know anything about them and start asking, "Oh, you know, what do you do? What kind of stuff do you work on?" And they'll have this like incredible project that they're using to build their rugby league and like recruit people for this league. And you're like, "I'm amazed that you came up with this," you know? So it's always cool to get to both run into people, but then also see them featured in the program itself is cool, too.

• 18:02Daniel Whitenack?

  Yeah, definitely. And I did look it up, uh, because you know, the internet can help with these things. And it is open source.org that at least is partially behind the conference. I don't know the exact structure there. But yeah, cool. Well, we're really excited that you're getting featured there and also a lot of other machine learning things are getting featured there that I, that I saw. But maybe just for our listeners, so probably a lot of our listeners that maybe come over from the changelog or another one of the changelog podcasts might be familiar with, with open source and kind of that community in general. But some like data scientists and AI people maybe coming from academia or working in research, maybe it's a little bit less clear, like what open source means

• 18:46Daniel Whitenack?

  and where you sort of like, how you get open source software and is it always free? What are the sort of, like, what is the community around it? How is it created? Could you kind of talk a little bit about that? Just kind of like, what is open source and how you were initially exposed to open source, maybe?

• 19:07Speaker 03

  Open source software is this crazy, amazing collective of people who see a problem with something or need to do something and somehow come together and self-organize to create a lot of the world's software. I am sure that if you're listening to this podcast, you are using open source software at some level, whether it's the server that's hosting this MP3 file or if it's the actual software you're using to play it back. I'm sure there's some part of open source software in there.

• 19:36Daniel Whitenack?

  Yeah, so it's not just like free apps or something. It's like, in a lot of cases, tooling or lower level things that are just available widely, that are integrated into all sorts of different applications. It may even be commercial applications.

• 19:51Speaker 03

  Yeah, absolutely. And a lot of the, like I was mentioning, web servers, a lot of web servers run on Linux, which is this amazing open source piece of software that people have been building and maintaining for decades now. And it's really interesting to see the kind of collaboration that the internet enables and the kind of governance models that people come up with. Like you said, not all open source software is free. There are some discussions within the open source and free software communities about the differences between those. And when people talk about free software, there's either the free as in you're not paying for it or free as in sort of like a free as in freedom or free as in liberty kind of thing. And some people really appreciate that aspect of it, too, because when you have software that is free as in freedom,

• 20:35Speaker 03

  that means that you have the right to change it. And if it doesn't suit your needs, you can go in and modify it. And there's some aspect of that that can be really satisfying for people who are using software and always want to have the ability to make it work like they want it to.

• 20:48Daniel Whitenack?

  Yeah, and we should just mention, kind of, for those just getting into software or AI or open source machine learning tooling and all of this, like, most of the time you can find these sorts of projects, for example, on GitHub. Not always. There's definitely like various NLP things and other things that I've used where it just like... Download it off of some university site or something like that. But for the most part, anyone can like create a GitHub repo and open source their like project or their the code associated with their research paper or whatever that is. I'm curious, Samuel, have you kind of open sourced things or been involved on that side of things?

• 21:31Speaker 03

  I have a little bit. There was a project that I worked on in college where I, at the time, couldn't find a good recommender systems library for Python and sort of hacked my own together and then put that onto PyPI. That is long since dead, but it was really informative, learning how to package up Python software and get it out to where people could actually use it. So I think even if the project doesn't become a success, like there's way better recommender systems libraries out there for Python now than that thing ever would have been. But despite that, there was still a lot of value in learning how to do that kind of stuff because you're gonna, you know, potentially need to ship something to your internal PyPI at work or something like that.

• 22:12Chris Benson?

  So, you know, it's interesting, unlike, you know, the original software development world where there was lots of closed source and open source kind of grew over the years and even over the decades. We've had this really cool situation where machine learning and AI tools have started off as open source. I mean, some of the most popular, you know, in this area are like TensorFlow, which was open sourced by Google and PyTorch, which was open sourced by Facebook. And so I guess my question is, why do you think that having open source in the machine learning and artificial intelligence fields is important? And, you know, why do you feel that they probably started it off that way?

• 22:54Speaker 03

  Yeah, so let's do that in two parts. I have a theory as to why things are this way. My, my personal theory is that really high level researchers, people who are advancing the state of the art, really like to be able to publish their work openly and be recognized for the cool work they're doing. So my theory is that because they come from this culture of academia where it is important to publish things publicly, that, well, then, you know, I'm gonna publish my code as well because that's part of the research that I did. So that's kind of my theory. It's not empirically validated in any way, but that's my theory on why that would be.

• 23:30Daniel Whitenack?

  Spoken like a true data scientist.

• 23:33Chris Benson?

  What do you think about just trying to get uptake on your tool? And by making it, for instance, back in 2015, if Google had not open sourced TensorFlow, maybe it would not have gotten such tremendous uptake as it did, and subsequently, Facebook as well. Do you think there's an intent of trying to capture mindshare in the community?

• 23:54Speaker 03

  Absolutely, yeah. I think that's really important to these large software companies. My understanding is that when Google wrote a bunch of these papers for what ended up becoming Hadoop, they kind of saw the Hadoop world rebuild, a lot of the internal tools they had. And then, it's good because you sort of build that mindshare of knowing what MapReduce is, for instance, but then when you're hiring someone and they come in, it's like, well, this is a little different from the MapReduce you're used to. And so if they just start right out the gate, open sourcing TensorFlow, then when they are hiring new data scientists, they're going to know exactly what TensorFlow is and already be using it. So I think there's a lot of benefit to it.

• 24:32Chris Benson?

  I agree. I think that you hit on the point. I remember that going around the community, that exact issue of Google kind of saying, wow, we kind of created this upfront, but since we didn't open source it, somebody kind of turned around and we had to react to them later on. And I think that also happened to some degree with containerization with them. And so maybe in that particular case, with that particular organization, maybe that was a lesson learned, that they finally turned it right. And obviously, that must have worked well for them because TensorFlow has a huge percentage of the market.

• 25:03Speaker 03

  Absolutely.

• 25:04Daniel Whitenack?

  Yeah, and I think that there's such a wide range of open source things now. I mean, I can't imagine doing any sort of AI project without open source tooling of some kind, but it's even past TensorFlow and PyTorch now, where people are kind of sharing their pre-trained models, they're sharing data sets, and all of those things. And of course, this is kind of also... Created some, like, a little bit of backlash in the sense that, like, you know, OpenAI models and others have been kind of deemed, like, dangerous and, oh, maybe we shouldn't be releasing this code. And there's also, like, it's kind of weird now

• 25:48Daniel Whitenack?

  that research is so close to applications. So, like, people release a paper and then there's, like, three implementations on GitHub, like, the next day, right? So what's your perspective on that, Samuel, in terms of, like, should researchers and data scientists always be open sourcing things? Or do you think that there's limits and boundaries within which we should work?

• 26:13Speaker 03

  I think, of course, there are times where it doesn't make sense to open source something or to release a data set. For instance, I imagine if you're Visa and you have a fraud detection algorithm, you definitely don't want to release that because then people are going to start doing really good credit card fraud to evade your algorithm, right? But there's other cases where I think it does sort of help the species, like, humanity, to get further along and understand how we can do certain things. So I will not pretend to be an expert about when is best. I can sort of see that there's cases where it's good and cases where it's bad. I probably tend toward thinking that there's something noble about trying to advance the frontier of human knowledge. But at times, that isn't the right choice. And at times, you have to have to make a difficult decision and not

• 26:58Speaker 03

  do that.

• 26:59Chris Benson?

  You know, before, I just thought I'd mention, I can give you an almost comical instance that I just read about where maybe open source wouldn't be a solution, which is strange coming out of my mouth because I think all three of us are huge open source advocates. There was a push by the Government Accountability Office for Department of Defense to, or kind of really all government agencies, to try to open source 20% of the code they write, which, which in general, I love that idea. But there was pushback from the DoD CIO, the Department of Defense CIO, and I'm paraphrasing, I'm not quoting because I don't have it in front of me. But he basically said, "Well, our code is building, is for weapon systems. And so we're not going to put the weapon system code out there in open

• 27:42Chris Benson?

  source," which I just, I thought was kind of hilarious personally.

• 27:47Daniel Whitenack?

  Yeah, I mean, there's definitely, I don't know, domain-specific situations like that. I guess there's like, there is this side of like, research where it's like, oh, we don't know exactly the implications of this, so maybe we're gonna, maybe we're gonna hold off. Anyway, moving past some of those like, caveats, I guess, what are some of the, if you were to pick some of your, your favorite open-source machine learning or AI projects that you use really frequently, what would, what would those be, Samuel?

• 28:20Speaker 03

  So I would have to start with Jupiter, because I think almost every data scientist has at least run into a Jupiter notebook at some point. I don't know that I've had a day without using one since I started working. They're just incredibly useful ways to sort of see the results of your computation and experiment with things and prototype with things in a way that can be a lot less friction than traditional IDE. That'd be what I would start with. Also, have to give a huge shout out to scikit-learn, because it has an incredible API, the community is really strong, the documentation is really good. And you can get a lot done with just everything built in scikit-learn, which is great. So those would be the two that I would give the biggest shout out to.

• 29:02Speaker 03

  If you have Jupiter and scikit, you can do a lot of stuff, for sure.

• 29:16Speaker 04

  Greetings, AI practitioners. Jared here, wanting to let you know that Changelog will be at All Things Open on October 14th and 15th. We're hosting a live JS party onstage, and as a special thanks from the organizers, we're giving away five free passes to the conference. All you have to do is tweet, "I want a free pass to All Things Open because", state your reason, and mention @changelog or @practicalaifm, so we see it, and we will DM you if you win. Okay, that's all for me. Let's get back into it.

• 29:54Chris Benson?

  So I am assuming that Indeed is not paying you to make music recommendation systems like you had mentioned in your ATO abstract. What kind of side projects are you engaged in? You know, what are you doing and what's fun? And what would you like to do that you may not have gotten to yet?

• 30:10Speaker 03

  Definitely. So there are machine learning related side projects and there are just sort of fun side projects. So as a fun one, I'll start with, I do some volunteer coding instruction and I think that's really enjoyable. And that's one of the most fun things that I do that's not work. As far as machine learning related projects, and these are some that I talk about in the talk actually, is working on some recommender system stuff I think has been really interesting. Trying to predict whether a certain musical artist will be liked by a certain user, I think is a hard problem for sure, but really interesting.

• 30:43Chris Benson?

  Are there like open source data sets related to that?

• 30:46Speaker 03

  There are, yeah. There's a conference called Rexis that released this massive data set that people gathered LastFM data. And LastFM is this social music sharing website where people, their music client will submit that they're listening to certain music. And then some researchers went out and scraped a bunch of data and put it into this thing. Now that's part of the open record and you can get this data, which is interesting.

• 31:09Chris Benson?

  Like, so you had that data in terms of like doing the recommendations. How did open source factor into kind of the way that you went about implementing a solution to that?

• 31:20Speaker 03

  Yeah, absolutely. The main thing I would say that it was helpful with for this specific problem was trying to do data exploration and visualization. I mentioned Jupiter earlier and that was a big part in the prototyping phase of this project. Another thing that was really useful was Pandas, which is a really good library for dealing with tabular data. If you have data that is rows and columns, use Pandas. It's great. And then also leveraging some tools like Matplotlib and Seaborn to do data visualization and try to see what sort of correlations exist in the data to try to get a first pass at what might be a useful model to start to build. I think those tools ended up being really useful.

• 32:02Chris Benson?

  Do you have any other projects in particular that you've worked on? Or anything that you're aspiring to when you get enough time to?

• 32:09Speaker 03

  Yeah, so there's one that I started working on actually before I got it indeed, where I wanted to use machine learning to find my next job. I just thought that was a really fun idea. And what that ended up being was I made a spreadsheet and read just a ton of job descriptions and pasted them in this spreadsheet and then rated them. And if I'm being honest, I definitely spent more time reading job descriptions this way than I would have any other way. But I would read them and try to figure out if they were cool or not and then have a training model to try to do this for me. And what I ended up doing, I mean, I don't have this email going anymore because I like my job, but what I ended up doing was having it do this weekly email where it would send

• 32:52Speaker 03

  me the top 10 jobs that sounded the coolest that went up that week, which I just thought was a fun little way to make your own life easier. I think that's a great way to get started with these machine learning projects. And like I was alluding to earlier, when you are trying to build a portfolio, it's cool to work on something that you actually want to solve because first off, it shows the potential employer what things you find interesting and you can inject some of your personality into that. And then second off, that you will be more motivated to work on the project.

• 33:21Daniel Whitenack?

  That's a great point. Yeah, definitely. I think passion is a big part of that that really helps with side projects. But yeah, I think you mentioned maybe a third project in your abstract having to do, was it something with sign language?

• 33:37Speaker 03

  Yes, yes. So there was a, actually at a hackathon, a friend and I built a little thing that would try to predict what sign you were making of the American Sign Language alphabet. So American Sign Language is a really interesting language that is not just English with your hands. It's much different from that. And a friend and I went to a hackathon and we had this little device that could read hand position data when you plugged it into your laptop. And we thought, what if we could like, teach a computer sign language? That sounds really awesome. Wouldn't that be cool? And then we realized we only had 24 hours and that was not going to work. But what we did realize was that we could at least start with trying to get it to learn the alphabet. And so, you know, we gathered some training data and then, you know, did some model selection

• 34:21Speaker 03

  and found a model that worked reasonably well. And now I get to tell people that I taught a computer sign language. It's not true. Like, the computer doesn't actually know sign language, but I can at least say that I, you know, with a caveat, was able to teach this computer something about sign language, which is cool. And what we ended up doing was turning this into a little learning game. We called it kind of like a Rosetta Stone, but for sign language, where it would show you a picture of a hand sign and say, hey, make an A. And you would make a little hand sign of an A above this sensor. And then, once you got it right, it would say, great job, and, you know, give you some points. So that was a fun hackathon project. It was really interesting and ended up having surprising applications to the real world of like, defining a limited scope and working iteratively.

• 35:03Speaker 03

  And ended up, I think, surprising myself with how useful that project and the things I learned on that ended up being.

• 35:08Daniel Whitenack?

  Yeah, I'm so, uh, I'm so happy for you to share that. So, obviously, as our listeners know, I'm very much involved in the world of AI and language, and especially minority languages, which include sign language. And, uh, yeah, I think if anybody want, out there wants a really cool, innovative and, you know, highly impactful project to work on, like, working on sign language tech is really interesting AI-wise. I know a couple of people that are working in that area and just doing amazing thing, like, things like processing video from like, three different directions and reconstructing hand motions in 3D and all sorts of amazing stuff. So I'm glad you were, uh, you're able to, to share that.

• 35:53Daniel Whitenack?

  That was at a hackathon. How much time and how often do you, are you doing side projects? I'm kind of curious about that.

• 36:02Speaker 03

  Absolutely. So since I've got some things working, like I'm married and I have a volunteering thing that I do weekly, I don't put a ton of time into side projects at this point, just because there's other things that I'm choosing to prioritize. But I think they were a really great way to scratch that itch for wanting to learn how to do machine learning by practice. And then I think partially, partially because of other time constraints and partially because of the fact that I get to scratch that itch actually at work now, I don't feel as much need to do that outside of work. But every now and again, something will come up where I'm just like, I need this, I need this thing, I'm gonna go build it.

• 36:38Chris Benson?

  So what do you think makes a good side project? A lot of people I know, you know, do that. We have what we do at work and then we all have our own little things. What do you think makes it worthwhile? What lends itself to being a great side project?

• 36:51Speaker 03

  The ones that I have found most fun to work on and the ones that other people seem to think are the coolest are things that are sort of tangible. So as an example, I have a website where you can go and type in a word and it'll try to make up puns based on that word. So if you type in Sam, for instance, into the thing and you click, give me some puns, it'll say like, ah, we're all in the Sam boat. Ah, ha, ha. You know, get really bad dad jokes like that. And I mean, it was a fun thing to work on.

• 37:21Daniel Whitenack?

  Sounds really useful for my life.

• 37:24Speaker 03

  Yeah. I mean, it's a fun thing to work on because it kind of has an element of humor and joy to it. And then part of what's enjoyable about working on a side project is showing it to people. And when you have something like that and you're able to get it out into the world, like, anyone can go to puns.samueltaylor.org and try out this thing. Anybody can. And you can show it to your friends and say, hey, type your name into this, see what comes up. And it's really fun. So I think that's part of it is the joy of getting to share it with other people is really fun. In other cases, I think side projects don't have to be broadly useful. If you can find something that is extremely useful to you, that's a definitely a great place to start. If you can have this job email, for instance, that

• 38:06Speaker 03

  was really applicable to me at one point in my life and it was really useful and the time was well spent and I learned a lot from it. And you might be surprised to find out that things that are extremely useful to you end up being useful to other people as well. So I'd say those are the first two things. And then the last one, which we've already talked about, is something that you just find interesting and you, you feel an itch to, to scratch that thing.

• 38:27Daniel Whitenack?

  So like, let's say you're working on a side project or in your, your main job, you have some sort of machine learning AI project going and you say, okay, well, I have this issue or I need to do X, Y, Z. I need to do recommendation or I need to parse this type of data or I need to scrape this type of data or I need to train this type of model. There's so much open source out there. How do you go about like finding the right tool for the right situation and also sort of validating, especially if you're doing this for your job, how do you go about kind of validating if this thing actually works as it's supposed to or will be stable for any period of time or robust in any sort of way?

• 39:11Daniel Whitenack?

  If you're just kind of grabbing stuff, you know, off of GitHub, what is your process for, for doing that? Cause I know I, I've learned certain things over time, but I'm, I'm curious what your perspective is.

• 39:23Speaker 03

  Definitely. So I'm going to cheat a little bit and say that the easiest way to do this is to find someone you trust and ask them about that. So if, if you have other data scientists that you work with and you can walk up to them and say, "Hey, I'm running into this issue. Do you know any packages that might be helpful?" And if they, they know something, hopefully they vetted it. So that's, that's the cheating way. If you can't do that, then I, one of the things that I actually tried to instill in the talk is when you run into a kind of data that you don't know how to represent, just Google it. Like search, "How do I do text with machine learning?" For instance, and you'll get a lot of results. And I mean, you might have to wade through a little bit and figure out what kind of seems to be the most popular way of doing that and figure out, "Oh, I'll,

• 40:05Speaker 03

  I'll use TFIDF," for instance. So the third kind of way, I would say if you're looking at projects on GitHub, and this is kind of still cheating, but if you, you can do some amount of validation by looking at how many stars something has and how many forks things have. If something has a lot of activity related to it, it's a good chance that it is well maintained.

• 40:25Daniel Whitenack?

  So could you describe a little bit more like how that activity is represented? Like on GitHub, for example, you mentioned stars, activity, like what, how is that activity kind of represented or what, what could you be looking for?

• 40:39Speaker 03

  Yeah, yeah, of course. So when, when you go to a GitHub repository in the, the top right-ish of the page, there'll be a thing that says stars and a thing that says forks and they'll have little numbers next to them. You might see a project that has 27 stars. What that means is that 27 individuals have landed on that page and thought, oh, this is cool. I'll bookmark that for later. And they click the little star button. That's all it means. It's not a vetting of the quality necessarily, but it is some amount of measurement or proxy for popularity. And generally, if something is more popular, you'll generally have more eyes on it and more people depending on it and people to run into bugs before

• 41:19Speaker 03

  you do and that kind of thing. So then activity, the, that comes more into this forking idea. On GitHub, you can do what's called forking a repository, which is you basically make a copy of it in your own space. So that way you can edit code and modify it and fix whatever bug you ran into, for instance. And you can, you can use those forks as a way of seeing how many people are actively working on this project.

• 41:45Daniel Whitenack?

  Yeah. Also, sometimes what I'll do is go to a repo and, you know, if I'm considering actually integrating it into a project, actually look at like the, the commit history. So like when you are working on a piece of code and you make a change, there's a, there's a commit that happens. And, you know, if the last one that happened was in like 2014 or something, probably less likely that, you know, the, the code is actually getting updated. Although, although that's not always a bad thing, right? If it's a simple package that doesn't need updating, you know, maybe that's something different. But yeah. So I guess we're kind of coming to the end of our chat here, but I know you mentioned like Jupiter and SciKit

• 42:27Daniel Whitenack?

  Learn are really great projects. Are there, are there any sort of other projects that you'd like to highlight as we kind of wrap up here that people might want to check out?

• 42:36Speaker 03

  Oh yeah. There's, I mean, it's, there's too many to name. One other that I will talk about that we haven't gotten to mention yet is Facebook has a library called Prophet, which is really good if you run into time series data. As it turns out, time series data has some odd particularities that often show up. And you can leverage a lot of people's knowledge who understand those particularities very well by using this package. And I think that's another one of those great things about open source software is that it often embeds the knowledge of a large group of people.

• 43:08Chris Benson?

  So I got one last question for you. I'd like to, I'm just kind of curious, and I ask people this all the time, is what choices in terms of, you know, what software, you know, obviously people are like TensorFlow versus PyTorch, you know, and others along the way, what kind of workflow and what tooling to support that have you chosen for your own personal workflow?

• 43:30Speaker 03

  Absolutely. When you're choosing between different implementations or different packages, my opinion is that the best thing to do is to make a prototype of something in both and make sure you understand what the benefits and disadvantages are of each one. If you can, like ship both to production, have them behind some feature flag or something, test them in some way, and try to see which one matches your use case better.

• 43:56Daniel Whitenack?

  Awesome. Well, thank you for sharing a little bit of your workflow, sharing a little bit about like what you're passionate about in open source. It's been really great to have you on the show. I know we've mentioned a lot of open source things in the past, but it's been really great to have someone kind of just share their perspective on open source and machine learning in general. And we hope you have a great time at All Things Open and, you know, hope that the talk goes well and you have a, have just an awesome experience there. Thank you for taking time to talk to us.

• 44:25Chris Benson?

  Thanks a lot.

• 44:26Speaker 03

  Yeah, of course. Thank you. I'm so glad that you enjoyed it and I had fun talking to y'all too. Thank you so much.

• 44:33Speaker 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 on 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 change log is provided by Fastly. Learn more at fastly.com. And we catch our airs before users do hear change log because of Robar. Check them out at robar.com/changelog. And we're hosted on Linode Cloud Servers. Head to linode.com/changelaw, check them out, support this show. This episode is hosted by Daniel Whitenack and Chris Benson. 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.

• 45:16Speaker 00

  Get our weekly email, 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.