From symbols to AI pair programmers 💻

How did we get from symbolic AI to deep learning models that help you write code (i.e., GitHub and OpenAI’s new Copilot)? That’s what Chris and Daniel discuss in this episode about the history and future of deep learning (with some help from an article recently published in ACM and written by the luminaries of deep learning). Sponsors: PSSC Labs https://pssclabs.com/practicalai – Solutions from PSSC Labs provide a cost effective, highly secure, and performance guarantee that organizations need to reach their AI and Machine Learning Goals. Learn more and and get a FREE consultation today at pssclabs.com/practicalai https://pssclabs.com/practicalaiSnowplow Analytics https://snowplowanalytics.com/ – The behavioral data management platform powering your data journey. Capture and process high-quality behavioral data from all your platforms and products and deliver that data to your cloud destination of choice. Get started and experience Snowplow data for yourself at snowplowanalytics.com https://snowplowanalytics.comFastly 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. Featuring: Chris 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: ACM article: “Deep Learning for AI” https://cacm.acm.org/magazines/2021/7/253464-deep-learning-for-ai/fulltextGitHub Copilot https://copilot.github.com/Books “Human-in-the-Loop Machine Learning” by Robert (Munro) Monarch https://www.manning.com/books/human-in-the-loop-machine-learning (use podpracticalAI19 for 40% off)“A Thousand Brains” by Jeff Hawkins https://www.amazon.com/dp/1541675797Upcoming 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.

• 0:00Chris Benson?

  I'm not terribly worried about whether or not a significant portion of code starts coming from AI sources in the years ahead, because if there's one thing I have figured out is that things are constantly evolving and we should stop trying to keep things exactly as they are today, you know, our comfort level right now. So whatever the future is, I'm going to dive into it. I'm going to enjoy it. I'm going to take advantage of that. I'm all for it.

• 0:27Speaker 01

  Big thanks to our partners, Leno, Fastly, and Launch Darkly. We love Leno. They keep it fast and simple. Check them out at leno.com/changelog. Our bandwidth is provided by Fastly. Learn more at fastly.com and get your feature flags powered by Launch Darkly. Get a demo at launchdarkly.com.

• 0:46Speaker 02

  With advancements in AI and deep learning evolving at lightning pace, it's more important now than ever to research the best options suited to your unique needs. This is particularly true when building custom systems and those systems that are GPU-heavy. Not only do the applications running on the system matter, but your AI infrastructure and budget constraints need to be front of mind as well. PSSC Labs, which is an HPC and AI custom solutions provider based in California, has been creating high-performance computing systems to meet their clients' unique enterprise computing challenges for more than 25 years. And with cloud computing costs growing at astronomical rates, plus companies increasingly

• 1:27Speaker 02

  losing control of their data security, it is no wonder that enterprises and government agencies need to continually look for ways to take back control of their data. Solutions from PSSC Labs provide a cost-effective, highly secure, and performance guaranteed that organizations need to reach their AI and machine learning goals. For more information and a free consultation, please visit pssclabs.com-practicalai. Once again, that's pssclabs.com-practicalai.

• 2:06Speaker 02

  Welcome to Practical AI, a weekly podcast that makes 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 changelog.com/community and follow us on Twitter. We are at Practical AI FM.

• 2:32Daniel Whitenack?

  Welcome to another fully connected episode of Practical AI where Chris and I keep you fully connected with everything that's happening in the AI community. We'll take some time to discuss the latest AI news and we'll dig a little bit into learning resources to help you level up your machine learning game. I'm Daniel Whitenack. I'm a data scientist with S I L International and I'm joined as always by my co-host, Chris Benson, who is a strategist at Lockheed Martin. How you doing, Chris?

• 3:02Chris Benson?

  I am doing good today. It's good to be recording and I think we have something pretty interesting to dive into today with everyone.

• 3:10Daniel Whitenack?

  Yeah, for sure. Well, I mean, in these episodes where it's just you and I, we kind of get the privilege to just talk about whatever seemed interesting to us. And there's been a lot. I know I've had a little bit of vacation and we've been in and out. So we've had some time to sort of scroll through, scroll through Twitter and the blog post world and all sorts of things to, to find a few things. The first thing that I wanted to talk about and mention was GitHub Copilot, which was trained by OpenAI. So it's a collaboration between OpenAI and GitHub.

• 3:49Daniel Whitenack?

  But essentially, the idea with GitHub Copilot, as I understand it, I haven't used it yet and I'm not on the team, but the idea being that it's sort of like an AI pair programmer where you have, while you're writing code,

• 4:07Daniel Whitenack?

  All of a sudden, you can sort of have the AI pair programmer step in and, you know, complete things for you or help structure your code, put in some things that it's expecting that you're going to need and that sort of thing.

• 4:20Chris Benson?

  Thank goodness. My god. My god, I've needed that for 30 years, man.

• 4:25Daniel Whitenack?

  Yeah, yeah. It's pretty clutch when you get that tab complete, but this is a whole nother level. So, like I say, I unfortunately haven't tried it yet. I'm on the waitlist. If you go to copilot github.com, you can sign up to be on the waitlist for it. So it's not everybody yet. And at least in my understanding, it plugs into VS Code, which I also don't use.

• 4:51Chris Benson?

  I do. This is perfect. Oh, I'm signing up right now.

• 4:54Daniel Whitenack?

  Maybe you could pass the waitlist, Chris, because one of the questions in the waitlist thing was like, how often do you use VS Code? And I chose the option that was like, you know, hardly ever or whatever, rarely, I think it was. So if there's any GitHub Copilot people out there, I would like to try it out. We'd also like to have you on the podcast, but I would love to try out the, uh, the system. But yeah, what are your general thoughts about this sort of thing in terms of, you know, how it's different from other developer tools in the past and what it means for us?

• 5:31Chris Benson?

  I think, I will say, you know, both of us have been programming for many years. And the thing about it is I will dip in and out of programming depending on what I'm doing and what projects I have going. And as I'm ramping back in at a given moment, I'm always trying to go, I know what I'm trying to do, but I can't remember the specific and that kind of thing. And that's where, you know, we both like Go, for instance. I know we met in the Go community. And I'll stop doing Go for a few months and then I'll want to do something and I'll swing back in and write some and I'll be like, ah, crap, I can't remember how to do this one thing.

• 6:05Daniel Whitenack?

  Yeah, it's like you're in Python world and then you want to put that colon.

• 6:08Chris Benson?

  Yeah.

• 6:09Daniel Whitenack?

  For your, after your for i in something, colon, and then you forget you're in Go without the braces.

• 6:15Chris Benson?

  Totally. This is going to help me ramp back in. And I'm sure it'll help me, you know, after that as well. But I've been waiting on this for my entire career, man.

• 6:25Daniel Whitenack?

  I mean, on the website, they talk about it. So it's powered by what they're calling Codex, which is a new, they say AI system. I assume that, I don't know if it's one or more models, but it seems to be consistent. And to be honest, this is all things I don't know because I'm not part of the team. I haven't interviewed any of them yet. Hopefully, we will be soon, but Exactly. I'm assuming it's similar to a, you know, these sort of large-scale language models of recent times that can like take in a lot of context and generate good language out of it. Although they say, you know, it's more than autocomplete in the sense that it's understanding

• 7:07Daniel Whitenack?

  both, you know, it's not just looking for an autocomplete for like the method on your object or something like that. It's actually looking at both your like your comments and your actual code structure, the function names, all of those sorts of things. It understands more of that structure within your code. Yep. More of the syntax. And so it can actually help you, help you and actually create some really interesting things. So all I've seen is other people using it and, you know, screen casts of people using it. And it seems like it does generate some really, really interesting stuff and some helpful things. So I hope I get past the waitlist. Me too.

• 7:50Daniel Whitenack?

  But I know that there were some people that were saying some things about, oh, is this sort of signaled the end? You know, now our AI is creating code and which could create AI, which could create code, which, you know, so that that's one thing scenario. The other thing is, you know, it does this sort of spell the end of software engineering as we, as we know it. What are your thoughts? I have my own thoughts on those two things, which you might guess from knowing me for a while, but.

• 8:19Chris Benson?

  Yeah, my thoughts, well, my thoughts, and this has changed over time. My thoughts have evolved even over the last few years as we've been doing the show. My thoughts are, I am totally, totally looking forward to taking advantage of this. I'm not terribly worried about whether or not a significant portion of code starts coming from AI sources in the years ahead. Because if there's one thing I have figured out is that things are constantly evolving. And we should stop trying to keep things exactly as they are today, you know, our comfort level right now. So whatever the future is, I'm going to dive into it. I'm going to enjoy it. I'm going to take advantage of that.

• 8:57Chris Benson?

  You know, there's the paradigm. I've recently learned, not everyone's aware of this, but there's kind of the idea, the meme of the lazy programmer. And that is, you know, you create great tools because in theory, because you're a lazy programmer, you don't want to keep doing the same stuff over and over again. So in the meme of the lazy programmer, I want this AI to help me so that I can get through my task and go relax and spend time with my family and chill. And produce amazing code along the way. I'm all for it.

• 9:26Daniel Whitenack?

  Yeah. What do you think are the dangers of something like that? I mean, they obviously released it in the sort of waitlist type thing.

• 9:33Chris Benson?

  Yep.

• 9:33Daniel Whitenack?

  Probably similar to like, you know, you remember when GPT-3 was released? For those that don't know, that's a very large language model. It's very good at generating like very realistic looking language. And there was a lot of talk about dangers around that and they released it in a restrictive way. What are your thoughts there?

• 9:52Chris Benson?

  I think that a tool like this is not a strategic threat to us. I think that this is a tactical goodie that we get to use to make our lives better. I'm going to worry a whole lot more when someday in the future, we're getting to AI capabilities that are able to think divergently of all the different things and figure out which one makes sense and what the long-term strategies are for those things. Because you know, that's what human brains are still really quite good at. And we don't have any AI that's remotely approaching that anywhere.

• 10:31Daniel Whitenack?

  Yeah.

• 10:32Chris Benson?

  It's very narrow in its capability right now. So I'm not worried at all about this. You know, someday I may worry about a future thing, but not this.

• 10:39Daniel Whitenack?

  The danger, if there is one, probably is in the fact, in expectations of this, right? If you expect too much of this, you could just start having this fill in a bunch of things. Of course, we know that the model's not going to perform 100%, right? And so that could introduce very bad breaking changes into your code base, which are also very difficult to debug, right? Because you didn't actually write every single line of code. So I think that it's not so much a danger of like job security or something like that, or, you know, sentience or singularity. It's more in the human computer interaction that happens when you introduce something like this

• 11:23Daniel Whitenack?

  into your development cycle.

• 11:26Chris Benson?

  So GitHub, folks, get it right, please. Get it right.

• 11:29Daniel Whitenack?

  Yeah, yeah.

• 11:29Chris Benson?

  Make this thing work.

• 11:30Daniel Whitenack?

  The other thing cool that I saw is it actually does, so it lets you write out comments and then it will sort of convert those comments into code. So you can sort of write comments like, I'm going to do this, and then I'm going to do this, and then I'm going to do that, and then it sort of generates the code for that, which is very different from autocomplete, right? It's more of like a, what's the right way to put it? It's more of like a machine translation type problem.

• 11:56Chris Benson?

  I love the idea. I'm going to create a development branch to go try an idea, and we can try that out. And if it turns out to not work, then I'll trash the branch and move back, you know, and try again or something. But I'm all about using it. I want them to take the pain of debugging away from me. I want them to fix that. That would be good.

• 12:14Daniel Whitenack?

  Yeah, the other interesting thing, I forget when, I listened to a talk once, and I really wish I could remember. Who gave the talk? If some of our audience remembers and they want to let me know where this talk happened in our Slack channel or LinkedIn or Twitter or somewhere. But there was a talk about GitHub, open source and where code lives. And basically, the thought with that was people underestimate how much code is living in private repositories. Yeah. That people don't have access to do these sorts of projects. Right? I'm sure that GitHub, I'm assuming, you know, maybe this can be answered somewhere on

• 12:58Daniel Whitenack?

  the internet, but I'm assuming that GitHub didn't give OpenAI all private repositories on GitHub to train their model. It seems like that would be a sort of breach of some type of agreement. So this was probably built on top of like, open source repositories, I'm assuming, of code. I'm pretty sure that they would record. If nothing else, their attorneys would be telling them. Yeah, yeah, yeah. Don't go down that other path. I'm sure there was a lawyer involved somewhere. There must have been a lawyer somewhere. But the result of that is basically that if you're building some type of expansive system like this on top of, quote, "all code," which is really not all code, it's the code you can get access to. Right. Which is public code. Right.

• 13:44Daniel Whitenack?

  It actually has a severe bias because there's so much code living in private repositories, because companies still primarily use private repositories, that actually you're sort of missing a bit of the goodness of like, code that is actually robust and supported. Yeah. And you're biased a little bit towards like, people's side projects that they post on GitHub and, you know, maybe don't actually work that well. They talked a little bit about bias and like, how the coverage of languages, like, does this work for Python and Go and JavaScript and whatever? So I totally get that, but I think also there's another relevant side of the bias

• 14:26Daniel Whitenack?

  here, which is, you know, what code is driving these suggestions? In my experience, you know, I would love to have that code be driven, or those suggestions and that generated stuff be driven by. Good quality code and biased towards that rather than sort of the bulk of code that's out there that's maybe not as good of quality. So I'm wondering if there's some type of filtering that went on to figure out, hey, what is the quality code to train this model on?

• 14:55Chris Benson?

  There are a lot of static tools for code quality. Maybe I'm hoping maybe they biased it by only selecting those that did very well in static code analysis and then use those as source.

• 15:06Daniel Whitenack?

  Yeah, yeah, I don't know. I hope that we can get someone from there on the show and probably some of these questions are answered in their frequently answered questions on their website, the Copilot website, which we'll point you to in our show notes.

• 15:19Chris Benson?

  There you go.

• 15:20Daniel Whitenack?

  So go ahead and check that out and let us know what you think.

• 15:33Speaker 01

  This episode is brought to you by Snowplow Analytics. Snowplow is the behavioral data management platform for data teams. Maximize the value of your behavioral data using Snowplow Insights, a managed data platform that's built on leading open source tech leveraged by tens of thousands of users. Capture and process high-quality behavioral data from all your platforms and your products and deliver that data to your cloud destination of choice. When marketing needs to make data-informed decisions, when product needs next-level understanding, and when analytics needs rich and accurate data, Snowplow is the solution for data teams who want to manage the collection, processing, and warehousing of data across all their platforms and products.

• 16:14Speaker 01

  Get started and experience Snowplow data for yourself at snowplowanalytics.com. Again, snowplowanalytics.com.

• 16:43Daniel Whitenack?

  Okay, Chris, you pointed me to an article. This must have been one of the ones that you saw in the past couple of weeks. I think it's a very recent article.

• 16:51Chris Benson?

  It is.

• 16:51Daniel Whitenack?

  That you wanted to talk through. And I thought it had an interesting concept in that it talks through a lot of the, you know, looking back on the progression of AI systems and how those have changed over time, leading up to deep learning and maybe what's beyond. And that's from Yan Le Cun, Geoffrey Hinton and Yoshio Bengio. And a recent ACM article in the communications of the ACM from July, from this month of 2021.

• 17:19Chris Benson?

  Yeah.

• 17:20Daniel Whitenack?

  Um, which we'll link that in the show notes for people to find it. But what, what struck you about this article that, that stood out to you?

• 17:26Chris Benson?

  Well, before we even dive into the article itself, it's notable that those three pioneers, I mean, they're often referred to as the godfathers of deep learning or the godfathers of AI. And that the three of them together received the Turing Award in 2018. You know, and these days just Bengio is at the Montreal Institute for Learning Algorithms. Geoffrey Hinton is at the University of Toronto as, and Google. And Yan Le Cun is the chief AI scientist at Facebook. You can't become a bigger luminary than these three in the AI world.

• 17:59Daniel Whitenack?

  Yeah. And the, for those that don't know the Turing Award, it's sort of like a Nobel Prize type award.

• 18:05Chris Benson?

  Yes.

• 18:06Daniel Whitenack?

  But in the sort of computing world.

• 18:08Chris Benson?

  It is huge and it is given out to these three. And this is being done by the Association for Computing Machinery, ACM. And if you're not familiar with that, it is one of the big giant original computing associations, you know, maybe, but you know, certainly among the most prestigious in the world, if not the most prestigious.

• 18:27Daniel Whitenack?

  Yeah, you'll see a lot of papers published in journals and, you know, events and other things.

• 18:33Chris Benson?

  And the title is just Deep Learning for AI. It's a very generic title, but.

• 18:39Daniel Whitenack?

  Yeah, when you sent it to me, I'm like, I don't. I don't know if this seems interesting from the title.

• 18:45Chris Benson?

  Yeah.

• 18:45Daniel Whitenack?

  But then I saw the authors and I'm like, you know, so you gotta read it.

• 18:49Chris Benson?

  It was interesting because of where they're going. So they're essentially, there's a lot of conversation as we record this. First of all, this is a very new article. This is from this month as we record this. We are in early July 2021 as we record this today. And this is a July 2021 article by these three luminaries all together for the communication and one of the biggest, you know, most important computing associations in the world. And so the fact that they chose to publish together means they're trying to make a point and they're standing together to do that. And if we want to foreshadow a little bit of where we're going on this, but maybe not go into the detail yet, they're talking about the future of deep learning and it finishes that way.

• 19:34Chris Benson?

  And they're talking about why they're still very bullish about, this is where the title comes in, about deep learning as AI, or in the case of the title, deep learning for AI. Because there's a lot of conversation going on right now in this industry about whether deep learning is the right paradigm for the future of AI, or should there be other approaches that do not fall within the realm of deep learning to carry us forward that might be more closely associated with how a human brain operates? Because there are very substantial differences in most deep learning models and how a human

• 20:16Chris Benson?

  brain is believed to operate and that research in neuroscience has been, has been developing. And it's something that I'm very interested in and I've been following. So I wanted to read this partly because I will admit from the very beginning, I'm a huge skeptic. I'm a giant skeptic and I am certainly do not have the qualifications that these three gentlemen have. I'm just your typical AI career professional out there and we do this podcast, but I'm struggling a little bit with that and that's one of the reasons I wanted to bring it up. And so I thought they start the article kind of doing a little bit of history. And we don't want to take everyone deeply, but they kind of hit the high points about why, why they're so bullish on deep learning for the future of AI.

• 21:01Chris Benson?

  And so they start off going all the way back to a section that's from hand-coded symbolic expressions to learn distributed representations. And, you know, Daniel, do you want to take a stab at, you know, kind of, most people probably in this industry don't really know what symbolic expressions even are?

• 21:19Daniel Whitenack?

  Yeah, I mean, to be honest, it's not really an area that I even have a lot of experience with. When I think, so my sort of introduction to, or when I first started hearing those words was when I was using the, some people recognize a computer program called Mathematica.

• 21:35Chris Benson?

  Yeah.

• 21:36Daniel Whitenack?

  Which is like, I think at some point, they, they were talking about, you know, symbolic computations and that sort of thing that are done on symbolic equations. But, uh, yeah.

• 21:47Chris Benson?

  I don't think we should actually spend much time on it because it's mostly thought by most people, there are exceptions, as something from the past that we should not return to. And there are going to be people that are just screaming at me right now that I just said that out loud. But I'm talking about if you survey, you know, a hundred people in the room, there might be two or three that still believe in that, but the majority of them aren't. And to give you a sense of this, since I'm very open about my age, I'm, I'm 50. And my mother and father, when I was a little kid, they were engaged in symbolic AI back in the 1980s at the Georgia Institute of Technology, Georgia Tech. And that is a pathway that has largely

• 22:29Chris Benson?

  fallen to the wayside. And so they kind of talk about, you know, going through that, they talk about some of the benefits of symbolic approach in the article. They also talk about why, why it moved on to deep learning. And then it's important to notice that, you know, people associate neural networks with deep learning, but neural networks predate the concept of deep learning by decades.

• 22:52Daniel Whitenack?

  Right.

• 22:53Chris Benson?

  They are older than I am even. And so there was already some advancement in a limited capacity in neural networks in the eighties and nineties, 1980s and 1990s. Then we had yet another nuclear winter and finally came about with the rise of, of deep learning and, you know.

• 23:09Daniel Whitenack?

  Or AI winter.

• 23:10Chris Benson?

  Yeah. And I'm just curious.

• 23:11Daniel Whitenack?

  Maybe nuclear winter too.

• 23:14Chris Benson?

  Exactly. For AI, it was a bit of that for a while. I mean, that's really where the term came from.

• 23:19Daniel Whitenack?

  Yeah.

• 23:20Chris Benson?

  Then the rise of deep learning came about.

• 23:22Daniel Whitenack?

  Yeah, I think I've talked about this on the show before, but in my PhD, the part of my original work was in what's called ab initio modeling in physics. That means you're just using the expressions and symbols to model a physical system. Even if you're doing sort of like computational integration or something like that, a numerical integration, you're still, you don't have fitted parameters. It was probably around the same time people started doing this very expansive fitting of parameters in our space for doing modeling of atoms and molecules. And I sort of view it as a similar process to that, where part of the reason why that

• 24:06Daniel Whitenack?

  became useful is also because of the emergence of certain computing, right? And they do talk in the article about the emergence of GPUs. And whereas in the past, you could have a very shallow neural network, meaning that, and for those that are maybe newer into this, when we're talking about shallow neural network, we mean that the neural network is composed of nodes. Each of those nodes takes in a series of inputs. They're added together with coefficients called weights and maybe another parameter called a bias. Those are added together and maybe another function is applied called an activation function. And that itself is, it's kind of just like input, output.

• 24:48Daniel Whitenack?

  It's like a function, but then you can start layering these functions together, right? And have like sub-functions of functions. And so shallow networks don't have that many of these layers. And so when they're talking about-

• 25:01Chris Benson?

  They usually have like one hidden layer in them.

• 25:03Daniel Whitenack?

  Yeah, exactly. So maybe an input and then a hidden layer of these sub-functions and then the output.

• 25:10Chris Benson?

  An internal layer, a hidden layer is not the input layer and not the output layer. It would be any other layer. And there used to be only one, really because it was limited by computation.

• 25:22Daniel Whitenack?

  You can already imagine, like, even if I have 10 of those functions and there's, you know, three or four parameters each in a very small type of scenario, that starts actually to become a lot, you know, a reasonable number of parameters to, especially if I have to fit those. So I'm doing some model fitting and I fit those on data, you know, depending on what computers you had available, that could take a while. But then with the emergence of GPUs, now people were saying, "Well, we're not limited by that compute anymore. Let's start making it deeper." So adding more and more layers. And one of the things that I thought was interesting was that they talk about, "Hey,

• 26:02Daniel Whitenack?

  it's partly that the deep learning had success because it was deep. There was the depth of these layers. It's really more than that." Let's see if I can find it here. The way they phrase it is that they exploit a particular form of composability in which features in one layer are combined in many different ways to create more abstract features in a next layer. And so you get more benefit than just like piling things together. You actually get a lot of sort of layered complexity that can be modeled.

• 26:32Chris Benson?

  It creates more capability for abstraction. So you get abstractions built on abstractions, built on abstractions. So really, the deeper it goes, the more capable the network is, the more accurate it becomes. And they talk about that all the way to the end of the article, as they talk about the fact that you can literally scale it in terms of depth and width and get more and more improvement in what you're trying to get. And that's what we've been doing for years. With GPUs really advancing a lot in recent years, we have seen the fact that we have gotten these giant performance increases, which has fueled this whole AI, you know,

• 27:13Chris Benson?

  movement, the modern movement in the 2000s and beyond. And so I think that's important is the question is, is that scalability alone, and there's other aspects that which we'll talk about in a second, but that scalability has driven the whole movement. As going back to my original point, that doesn't mean it's exactly like what a human brain looks like is, as you know, does it matter? I guess that's what they would be suggesting.

• 27:37Daniel Whitenack?

  Yeah, yeah, I, I don't know. The whole human brain thing, I guess I don't think about too much. I think more about the practicalities of does this solve a task or a problem? Maybe it's because I don't know that much about human brains myself, that I'm not really that concerned of how well this models a human brain. I'm more concerned with using my human brain to apply these tools to solve tasks. Yeah, I don't know if you have any thoughts in terms of, is it even useful to wonder how closely these things model the human brain?

• 28:11Chris Benson?

  So I think that's where we're going with the argument is, I think that's the question. It's not how close it is, it's does that matter in the long run as we explore this? Because no one who's in this industry can deny the incredible utility of neural networks, of deep neural networks and all of these architectures, which we can talk about in a moment along the way. But it's still not the same thing. And we don't yet have human brain level capabilities. And most people would say we're not close to that yet. So the question is, to get close to it, do we, can we do it along this path? Or do we need a completely separate path?

• 28:51Chris Benson?

  So as we, as we look at the question going forward, you know, let's take a brief moment and just talk about some of the high levels of, of where things have gone. We've gotten different architectures that had been evolved out of the vanilla neural networks, such as generative adversarial networks. And we have natural language processing is using all sorts of deep learning capability. We have computer vision using convolutional neural networks and capsule networks and all of these other capabilities. In the NLP side, we've had a whole bunch of episodes on transformers, which maybe, could you take a moment? I know you work with them a lot and kind of explain what transformers are, because they've had a huge impact on, on where NLP has gone in recent years.

• 29:32Daniel Whitenack?

  Yeah, I think maybe it's it's twofold. Part of the utility of these large transformer models is in the fact that they utilize this unsupervised pre-training technique, which they talk about in the article. And part of it is the actual, you know, architecture that's used. And they also talk about convolutional architectures and where those are useful for other things. But the unsupervised pre-training bit, I think that idea goes beyond transformers in that the great thing that was discovered was you could create these reusable, almost modules or components that took your input data and extracted out features that were

• 30:16Daniel Whitenack?

  relevant for particular kinds of problems. So let's say that we're doing an object, you know, some type of image-based modeling, like object detection or image classification or something like that. If you train one model with a whole bunch of layers, which are typically composed of convolutional layers and maybe a very small number of layers at the end that do the actual task-related thing, like the classification, then what you end up is with a network or a model where the bulk of that model, the work of that model, is taking the input data and creating this set of features or extracting this set of features from the input data.

• 30:58Daniel Whitenack?

  It's only that last little bit that's used for the classification. And so what this ends up creating is if you chop off that last bit of the model, the part that does the task that you want, in the case of image-based models, maybe that's classification or recognition. In the case of text-based models, maybe that's sentiment analysis or named entity recognition or something. What you end up with is this pre-trained model that you can actually transfer to a whole variety of tasks. And I think that that's the big, you know, one of the big boosts in recent times is, you know, someone like Google or a large player in the space will release this large pre-trained model that is general.

• 31:42Daniel Whitenack?

  To a wide variety of problems in a certain space. Maybe it's text problems, or maybe it's image-based problems, or maybe it's speech-based problems. And this allows you to be much more efficient in terms of the actual usage of AI techniques, because you sort of stand on the shoulders of giants in that way. And so they do specifically call out this method of unsupervised pre-training as something that's driven a lot of success.

• 32:10Chris Benson?

  And we've seen more and more of that in recent years. I mean, when you and I started this podcast about three years ago, supervised training was still by far the dominant mechanism. And we have watched this over this time period. We have watched the steady progression into unsupervised training approaches, because A, it scales much better than supervised training does. If you have to have supervised training, meaning that your training data has labels associated with them about what is right and wrong. And as you train your network, you are going off those labels. That means you got to label everything. We used to have episodes on that early on, about the problems and the challenges. How do you do that? And the science of this has evolved over time to recognize the value of unsupervised.

• 32:56Chris Benson?

  Plus, when you think about it, a lot of, you know, just to kind of keep our eye on this kind of comparison with human brain, we do a lot of unsupervised training of different types in our own brain. And so that is, and often do it without many iterations as well. So as we are learning more about deep learning and applying these new techniques, we are, to some degree, taking, even where they're not the same, we're taking inspiration from the human brain and recognizing that unsupervised training is where most of the future is going to be.

• 33:27Daniel Whitenack?

  Yeah, and coming back to the sort of transformer thing, which people might most commonly associate with NLP or language-based modeling, has actually turned out to be a very significant advance because It's applicable beyond that as well, because in some ways, it's like moving from a sequential or, you know, very strictly connected type of network into something that's very dynamic. And so how that happens is like this, in sort of traditional ways of processing text, you see text as a sequence, right? So there's the first word, and the second word, and the third word, and the fourth word.

• 34:12Daniel Whitenack?

  And so that's when people started thinking about recurrent layers, which process things in a sequence. Now, you could have bidirectional layers, so there's like bidirectional LSTM layers, which are very popular, and that looks at your sequence of inputs from left to right and right to left, and you can sort of do really, you know, creative things like that in terms of how your model is actually processing the data that goes through it. However, like, as your sequence gets longer and longer, and as your space of possible inputs, like your vocabulary, gets larger and larger, which is what was happening with these larger and larger language models, right?

• 34:52Daniel Whitenack?

  We had tons of data they're training on, very large vocabularies, maybe they're looking at larger chunks of text to process. What they realized is this sort of sequential processing, I mean, one thing is, it's just not that computationally favorable. If you start looking at all of the sort of spaghetti of connections that happens as you kind of add in BiLSTM or recurrent layers, it gets very complicated. And so what happened was, partly out of computational reasons, people started looking at these layers called attention, or self-attention, or soft-attention layers, where actually

• 35:34Daniel Whitenack?

  you take your set of inputs, and you shove them all at once into a layer, and let the model decide which ones of the inputs are important to pay attention to as it processes them through. You still have this sort of sequence of things, and in some cases, you want to also put in something indicative of the sequence order. But in general, you can sort of, the model can decide on the fly which of the inputs are important to pay attention to. And if you start stacking these into what they call multi-headed attention, then you sort of have a whole bunch of combinations of different inputs and secondary inputs that the model can decide to pay attention to,

• 36:17Daniel Whitenack?

  even if, let's say, it's paying attention to a word that happened much earlier in the sequence, but it knows that it's really important to process that with words that occurred much later in the sequence. And so that sort of had a boom for NLP processing, both computationally and because it had this attention mechanism. But after that, this sort of attention mechanism and transformer architecture has been applied in a whole bunch of different places with images and with speech and all sorts of things because it applies this way of looking at inputs in a very dynamic way,

• 37:00Daniel Whitenack?

  which is actually significantly different from some of the layers that have been used for a long time, like convolutional and recurrent.

• 37:07Chris Benson?

  Yeah, and by the way, that was a great explanation of transformers.

• 37:12Daniel Whitenack?

  Well, it's not a full explanation, but there are really good explanations out there that I can put in the show notes. But I think it's that attention mechanism is really one of the key things. And they call out that attention mechanism in the article.

• 37:24Chris Benson?

  And then they get into, and they cover a few other things, some of the architectures. But with the time left, you know, then they get into the future of deep learning and kind of talk about, you know, where to go. And they start that conversation, kind of going back to that scalability thing that we talked about at the beginning of this. And that was, they specifically call out the language model GPT-3, which we have talked quite a bit about over time. It has 175 billion parameters. And they point out that that's still tiny compared to the synapses in the, in a human brain, but that it generates notably better text than GPT-2 that only has, only 1.5 billion parameters. And they also talk about some chatbots, but they're

• 38:06Chris Benson?

  Point is that you're getting better performance simply by scaling up in these areas. And then they go on to really talk about areas for improvement. And the three things that I'd like to throw out for us to talk about that they mentioned is they say supervised learning requires too much labeled data and model free reinforcement learning requires too many trials. And that we, which I think is a really good point here, humans seem to be able to generalize well with far less experience, which we talked about a few minutes ago, is that you and I can go decide to learn something new, depending on the complexity of the thing, and we might pick it up almost immediately, you know, without having to go through thousands of iterations.

• 38:47Chris Benson?

  They also mentioned that current systems are not as robust to changes as humans are because we, we humans adapt so quickly to changes with very few examples, which is more of the same.

• 38:57Daniel Whitenack?

  Yeah, I think that that second one is definitely a big one in my mind.

• 39:01Chris Benson?

  It is.

• 39:02Daniel Whitenack?

  Of robustness.

• 39:03Chris Benson?

  Me too.

• 39:03Daniel Whitenack?

  Because that both helps protect yourself in terms of your model behaving well, even with unexpected data, but it also helps in terms of, you know, you being able to utilize other people's models that might be adaptable with very few examples to new scenarios. And there's been a lot of work with this. I know we had an episode once where we talked through the OpenAI effort where they had the robot hand that was moving around the Rubik's Cube.

• 39:31Chris Benson?

  Yeah.

• 39:31Daniel Whitenack?

  And they sort of perturbed that system with a variety of things like a stuffed giraffe and they hit the hand and other things. And that was in the context of reinforcement learning and also some sort of multimodal inputs. And they had various ways there where they specifically called out their efforts to improve the robustness of the system. And I think that that's one area that I probably, I think about how well my models are doing on the test sets that I've organized. And I also, you know, I try to look at, hey, how is my model doing over time? Is it drifting, you know, in production? Those are all good things to think about.

• 40:14Daniel Whitenack?

  But this idea of like how to educate people on how to make their models more robust and maybe use adversarial examples and other things, that's one area that practitioners could in particular really grow in.

• 40:29Chris Benson?

  That's true. And by the way, the third item they talked about, even if we move on, is they have this idea of system one tasks versus system two tasks. And basically, kind of the simpler things that we are starting to do with deep learning are called system one, things like perception and, you know, these kind of basic level building blocks for putting together things. But there is the concept of system two tasks, which are more complex and things that, you know, in short, we might often associate still with human brain and less with deep learning. It's more aspirational on the deep learning side. And they acknowledge that we're in the infancy on that. And that's, just wanted to point out what the third one was. They kind of go into what needs to be improved. And the robustness is one of those areas that

• 41:11Chris Benson?

  you just talked about, that they really address upfront. And they talk about kind of how to do that. Do you have any, any more comments based on that? Or?

• 41:19Daniel Whitenack?

  I don't think so. I, I mean, I'm drawn back to the conversation we had with the developer of Text Attack, which is, was a package where he was trying to bring this idea of robustness and adversarial examples to tech-based problems where people don't really think about that as much. They might think about it more in robotics and perception type problems.

• 41:45Chris Benson?

  That was a good conversation.

• 41:46Daniel Whitenack?

  That was a good conversation. We'll, we'll link it in the show notes. But I think that my thought was like, how do I start integrating my thought process around robustness and adversarial examples into my own workflow? I think that's what I'm trying to, because often I'm just so focused on kind of getting through the problem, getting, you know, reaching a certain level of performance. And yes, I do want to go back and do updates to the model and see how it's performing. But it's not really like before I released the model, I've always done a ton of work around, you know, robustness and probing the behavior of my model, doing a lot of sort of minimum viability tests and perturbations to understand how it behaves. And maybe that's

• 42:31Daniel Whitenack?

  partly just the time that's allotted to me to solve a problem. That has to be balanced, I know.

• 42:37Chris Benson?

  We're almost to the end here of their article. The next thing they have is one that I'm keenly interested in. And that is kind of alternative architectures. So, you know, we took, you know, you described what a basic neural network kind of looks like and how the neuron operates earlier in this conversation. And I think it's key to recognize that those hidden layers that you that you were discussing are kind of homogeneous. They create these layers of abstraction by having many, many layers in there. Thus the word "deep," but they are homogeneous in the way that they operate. And then we have the other, all these other architectures that have evolved out of that are where people have something in mind and they make a change to the architecture.

• 43:21Chris Benson?

  They may introduce, you know, the recurrent pathways or there's, there's a whole bunch of different of techniques that they have introduced. But it doesn't fundamentally change that kind of baseline architecture there. And one of the things in brain research these days is the idea of having groups of neurons that represent entities. So they're not homogeneous. You have regions of neurons and synapses that they start as homogeneous. They start as a blank slate, but as the brain develops, they take on certain roles and they represent entities and the relationships between entities and those. And there's now a fair amount of work that's being done to kind of connect artificial

• 44:04Chris Benson?

  intelligence, you know, and I'm almost separating those, meaning that the neural network approaches back to these other possible architectures. And so there is a lot of discussion. I'm quite fascinated with the way some of those are going. And without foreshadowing, we may have some guests on before long that will address some of those because it may be a turn along the pathway. And I frankly think that that specific issue is one of the reasons that these three gentlemen wrote this article. At this point in time, I think, because I really think that they are saying, no, no, this pathway that we're on is the right path as opposed to making a turn or taking an alternate path along the way. So what do you think about that? What do you think

• 44:46Chris Benson?

  about this homogeneous, you know, generating these networks as we have been doing out of this homogeneous hidden layer collection versus considering other options for neurons being grouped together to represent entities and relationships?

• 45:00Daniel Whitenack?

  I think it's consistent with other conversations we had. I remember a brief comment from William Falcon, who was creator of PyTorch Lightning on his episode with us. He talked a little bit about how their PyTorch Lightning is structured in such a way so that you can do research into these sort of composable pieces and modules. Yeah, I'm excited to see what happens there. I think there will be additionally really interesting advances in new architectures that process data in interesting patterns or structures like graph neural networks that actually operate natively on different structures of

• 45:40Daniel Whitenack?

  data rather than, let's say, just, you know, matrices or tabular data. I think that's another interesting area in addition to this sort of groupings of neurons and all of that. But yeah, I'm excited to see where things go. I definitely think that for one, we have some exciting things to look forward to and for two, GitHub Copilot isn't going to write all our code, so we'll get to code some of it anyway. So I'm excited about that. I know in these episodes, we normally like to share a learning resource. I just want to just share something real quick before we end. I noticed this new book that came out in June called Human in the Loop Machine Learning by Robert Monarch. It seems to be

• 46:24Daniel Whitenack?

  getting quite good reception and I know I would like to go through it. It is from Manning, which is great because it means that we have a discount code from Manning for our listeners. So if you want 40% off, that's crazy, 40% off.

• 46:42Chris Benson?

  That's pretty good. That's a good discount.

• 46:43Daniel Whitenack?

  It's pretty, it's pretty good. You can use the code PODPRACTICALAI19.

• 46:49Daniel Whitenack?

  That's that book and we'll put that information in the show notes as well.

• 46:53Chris Benson?

  That sounds like a good one. I'll offer one as well, which I don't think you'll be surprised at when I mention this. There's a book that came out just a few months ago. It's on this topic that we're just finishing up with. It's called A Thousand Brains: A New Theory of Intelligence, and it's by Jeff Hawkins. And I have that book. It's a very interesting book, and I definitely recommend if you're wanting to explore some new ideas about where AI could go. It's definitely a book that will stimulate some thought. So that's my recommendation. A Thousand Brains by Jeff Hawkins.

• 47:27Daniel Whitenack?

  Awesome. Well, I'll let you get to your reading, Chris.

• 47:31Chris Benson?

  Absolutely.

• 47:32Daniel Whitenack?

  Thanks for the discussion.

• 47:33Chris Benson?

  It was a good one today. Thanks, Daniel.

• 47:38Speaker 02

  Thank you for listening to Practical AI. We appreciate your time and your attention. If you enjoyed this episode, help us out by spreading the word. Think of a friend, think of a colleague, somebody who would benefit from listening to it, and send them a link. We'd really appreciate it. Practical AI is hosted by Chris Benson and Daniel Whiteknack. It's produced by Jared Santo with music by Breakmaster Cylinder. Thanks again to our sponsors, Fastly, Linode, and Launch Darkly. That's our show. We hope you enjoyed it, and we'll talk to you again next week.