Not Your Father’s Data Center - Unpacking AI and Machine Learning

Episode Date: April 5, 2021

AI and machine learning are common terms, but they are complex with many applications. Understanding this is critical, and Not Your Father’s Data Center Podcast is back to unpack the terms with guest Zachary Lipton. Lipton is an Assistant Professor of Operations Research and Machine Learning at Carnegie Mellon, where he runs the Approximately Correct Machine Intelligence Lab.

Welcome to Not Your Father's Data Center podcast, brought to you by Compass Data Centers. We build for what's next. Now, here's your host, Raymond Hawkins. Welcome again. Let's see, today it is February 23rd. We are recording. We are approaching a year and a month into a global pandemic. Welcome again to another edition of Not Your Father's Data Center. I'm Raymond Hawkins, Chief Revenue Officer at Compass Data Centers, and I am joined by Zachary Lipton, who is the BP Junior Chair, Assistant Professor of Operations Research and Machine Learning at Carnegie Mellon University. So translation for our listeners, that means he's a lot smarter than I am.

Zachary, thanks for joining us. Thanks for having me. So Zachary, we'll dive in if you're willing, just to do a little bit of background on you. Certainly, it's a pretty technology-centric and data center-centric audience that we speak to, but found that for them, understanding and learning a little bit about who's going to be talking to us today is interesting and engaging and makes it far more fascinating for folks. So do you mind giving us a little bit of your history, where you're from? Certainly have your bio that you got a degree in economics and mathematics from Columbia and then master's and

PhD from UC San Diego. So we can have a great Keynesian and Austrian economics conversation on another podcast. But for this one, we'll stick to machine learning and AI. But tell us a little about you. Yeah, sure. I mean, in a nutshell, I grew up in New York, kind of sleepy suburb of New York City. I ended up going, I guess my first kind of passion was music. So I was playing jazz music and it was a cool time growing up in New York and not during a global pandemic. And while a lot of the great musicians were alive, like Brantford Marsalis lived in my town and actually his son and I went to high school together and I used used to go over to his house and get saxophone lessons and oh wow yeah

my main love like what I thought I wanted to do um and I ended up going to college at Columbia I kind of had like uh I don't know like I I really liked music but I also I I guess like maybe part of it was like the school of like learning jazz that I came from was like sort of you learn it from the community, not from necessarily from a school. So I kind of felt like I wanted to go to university to get like more of a secular education. So I ended up going to Columbia for undergrad and studying math and economics and then was also playing music. And that was kind of like an interesting balance for me. It was cool because I was in New York City and I could go out and play, but I also was kind of like an interesting balance for me it was cool because i was in new york city and i could go out and play but i also was kind of learning technical things

and had this sort of you know weird like different daytime life and nighttime life and and that kind of eclectic existence and i felt you know maybe balanced or right and then after i graduated from undergrad i was just playing music mostly for a while. Also had some, you know, help things that kind of derailed me for a little bit. And then when I was kind of getting back on my feet, it was sort of like, do you just go back to doing what you were doing before? And do you just kind of like pick a new adventure? And at the time, I had a really good friend who was doing a PhD actually in music. He was a great pianist and

he was studying composition at uc santa cruz so you know i had just been kind of you know like new york's not a great place to like have no money and be kind of sick and so like it was kind of like living in a like rent stabilized apartment where like i think the landlord's strategy was to like get everyone to move out by making it as unlivable as possible so they could eventually jack the rents up so it's like kind of in this like smelly like urine and you know like kind of building in new york and then i just like go out to santa cruz um and i had already been kind of inclining towards towards academia as like maybe the last place that that things kind of felt right and i went out you know from from being in new york where it things kind of felt right. And I went out, you know, from being in New York where it's kind of coming out of the winter

and it's like smelly apartment and then visited my friend at UC Santa Cruz, which is just like paradise, you know, like all the, you know, the 80 year olds look like they're 30 years old, like all the produce is like ripped fresh out of the ground. You know, you have a beautiful view of the ocean. And I knew I didn't want to do a PhD in music. But something about like the experience of doing a PhD, of being at

the university and specifically of like being in California on the coast, just like everything just kind of felt right. And so I just came back to New York, broke my lease, moved to California. And I was trying to figure out what I wanted to do a PhD in. It was sort of like the decision to go to grad school preceded the decision of what the field would be. And I was kind of mulling it over, and I felt like maybe I'd want to do something in life sciences. And so I was exploring that. And at some point, I was like,

I had sort of self-taught some amount of computer science while I was out of undergrad. And that had been something that clicked with me earlier. And it seemed like kind of a ridiculous idea that someone would let me into a PhD because I really hadn't programmed in like seven years and didn't really know much. But I don't know. I just kind of like this kind of idea of going to PhD just kind of snowballed and became sort of a weird fantasy that it just kind of like forced into existence. I basically came back, took the GREs, broke my lease in New York, drove across the country, moved to California, got really lucky that someone found, I guess, you know, found my application like weird enough or interesting enough to give me a chance at UCSD. And I entered the field of machine learning as my chosen area.

Well, I got a thousand questions I want to ask you. So this is a great intro. Thank you for giving us a little background and where you're from, your passions. Love hearing the jazz musician passion. We'll love to get your thoughts on the movie La La Land at some point. I don't know whether to love the movie or hate it,

but would love to ask you some specific questions around things that you do today. Before we do that, I'm going to do trivia question number one. This, again, you email us at the show with your answers. Everybody gets correct answers drawing for $500 Amazon gift card. Question number one, Zachary teaches at Carnegie Mellon University. Can you tell us who the university is named after and what did he do to gain notoriety? So that's trivia question number one.

All right, Zachary. So you run the Approximately Correct Machine Intelligence Lab. I'd love to understand the name Approximately Correct. I know it's also your blog as well. So give us a little history behind what do you mean by Approximately Correct and help us understand that a little. The name is a little bit of like a play on a few things. So one thing is in learning theory, the kind of like one of the like canonical framings of learning problems is this probably approximately correct learning. It's like, what can you say about, say, some predictive model, given that

you train on some samples? And you could say, well, with high probability, it is close to, you know, the optimal predictor or something like that. So it's like, you can't exactly solve for like the exact thing you are after. You're constrained by the fact that you're learning from data, but you can produce something that is approximately correct. And with higher probability, you can be closer to the right answer as you get more and more data. So one play, that's something that's sort of familiar to any machine learning academic is that usage of it. And we do have, I'm not at core a theorist, but I do have some students that are much further on the theoretical side and we do some ML theory.

So there is that kind of aspect of the play. The other side is the way just machine learning is sort of always used in the real world is as a sort of, you know, not quite right, but sort of gets the job done. And maybe the benefits of scale outweigh the ways you've mis-framed the problem. So it's sort of given, you know, in this other way, not in the learning theoretic sense, but in this fuzzier way, like machine learning is very often this sort of approximately correct solution for the kinds of problems

that we're directing it at. And then maybe like the third usage for me that was in mind when I started that blog. So I started that blog around the time that I was working on Microsoft Research. And part of it was a sort of annoyance with, you know, I had very strong opinions about like kind of public writing about science and how we communicate what's going on in research to the broader public. And one thing that always bothered me a lot was the way that people didn't seem to, people didn't

seem to communicate sort of thoughtfully or honestly about uncertainty. Like what are the things that we, you know, what are the things we know? What are the things we don't know? What precisely is uncertain? Or what are the holes in our understanding? And I feel like this was like an important part that was like missing from the picture

of a lot of writing about science generally and something I was trying to accomplish with my way of like relating to the broader public. So I think like those three things together, but you know, were the inspiration for the name. Will you help me understand the difference between artificial intelligence and machine learning? I know in my industry in the data center, we're excited about both because they drive more capacity in our space. But I don't think the folks that are in my business and largely in our audience understand the difference between the two. Can you give us two minutes on that, Zachary? That would be awesome.

My sense is that as far as the folks in your industry are concerned, so there's like a practical answer to what these terms mean historically and like semantically and the ways in which they're different. And then there's how are they being used in industry by folks today and what are people referring to? And so the short answer is I think as far as people like thinking about cloud compute and, you know, like, you know, using GPU instances to do AI or to do machine learning, the terms are used actually, I think, interchangeably. What's happening right now is that basically there's this way that people just brand them.

It's like some group of people have come up with some algorithms, some effective ways of using data to do whatever, ranking or extracting marketing insights or whatever, whatever, whatever. And they call themselves big data companies and everybody calls themselves a big data company. They say, oh no, we marketing insights or whatever, whatever, whatever. And they call themselves big data companies. And everybody calls themselves a big data company. They say, oh, no, we're not a big data company.

We're an analytics company. And then everyone calls themselves an analytics company. We're not an analytics company. We're a machine learning company, right? And so, okay, now some of those things are distinctions, but you can't like divorce yourself in the fact that a huge part of what's going on is that people are just are using the nomenclature as a way of like differentiating themselves so like you know it's like i sometimes get this example of like to show how it's kind of clownish is like imagine like physicists

you know like this had some breakthrough within a lot of people get interested in physics and then people say what are you working on he said oh no we're not doing physics we're doing schmizzix or something you know yeah no 100 no, 100% agree. Absolutely agree. But I think there's such a, like such an incredible amount of ignorance here that nobody knows what any of these companies are doing. They just know it involves data

that companies do feel this pressure because they aren't necessarily often dealing with like a customer on the other side who really knows what they're doing, such that just like the technology stands on its own. they feel this need for like perpetual rebranding. So I think like a lot of what's being called like AI was sort of a taboo word because it had a bad reputation. It was associated with an academic field that lost a little cachet, mostly owing to a feeling that had sort of like over-promised and under-delivered and had

maybe like you know been a little bit had over claimed a lot had not been so rigorous the sub field of ai that like you know got more focused on a very statistical way of doing things and specifically fitting models from data whatever uh branded themselves as machine learning kind of shook off the AI term. But as soon as it became super successful and super popular, mostly, you know, in the last, say, 10 years, owing to the rise of deep learning, which is just a specific class of algorithms within machine learning that are based on neural networks, suddenly, as it started getting popular again, then people started adopting the AI term. But, you know, whereas the adoption of the word ML versus AI

coincided with actually a change in, you know, like those people were casting off a whole family of approaches and types of algorithms and stuff they weren't interested in focusing specifically on this sort of statistical machine learning that actually coincided. So like with a real change in direction, at least among the sub community, the shift back to adopting the term AI is, you know,

as you're seeing it now describing broadly, just companies that are just doing machine learning, that to me is just, it's just marketing fluff. Now that said, I would step back and just point out that the two terms are a bit different historically. How deep you want to go down that rabbit hole? You know, maybe in two ways.

One is that the term AI originally meant it was a term embraced by the people who were sort of adopting this sort of logic-based or symbolic logic-based, you know, approach to building intelligent systems. Whereas machine learning really grew out of like a rival sort of approach to building intelligent systems. Whereas machine learning really grew out of like a rival sort of approach to building intelligent systems that was associated with a school of academics that were called the cyberneticists. And that might sound kind of antiquated or goofy, but actually the stuff that is successful now in machine learning, whether it's neural networks or reinforcement learning, or even like neural networks or reinforcement learning or even

like control theory, these things actually are the intellectual legacy of the cyberneticist, not of the people who created the term AI. So that's like the historical difference. I'd say that though more recently, AI has become kind of an umbrella term. And so like what I think is like the common usage among academics, at least, of how they sort of break these things apart is AI has become this sort of umbrella term

for sort of the wide discipline of trying to build technical systems that sort of perform tasks that until recently we thought like required human intelligence. So that may or may not involve learning from data, right? So like, you know, at one point in time, it was all about search, efficient search algorithms and tree search, things like the way they built Deep Blue to play Garry Kasparov at chess. There was no data involved in that.

That was all about, you know, efficient algorithms for tree search. Then within that is a specific family of algorithms we call machine learning. And machine learning is about algorithms that learn, and by which we mean learn, we mean that they improve at some task as a function of experience as they acquire more data. So the AI is about broadly any approach that like does things that we think require human intelligence. Machine learning is specifically about learning from data. So in that view, which I think is the most productive and maybe closest to universally held among actual academics now of the most useful deployment of those terms, like AI, you could think of as like a wider tent and ML as like a narrow subset within it

that's specifically focused on things about learning from data or learning from experience. And I say, it just so happens that most of the action in the last 10 years, most of the real change has been concerning the machine learning subset. You know, to the extent that we're suddenly now renaming it AI, I think it's, this is

speaking more to just like a need to like keep the brand fresh or something for marketers to say, well, you know, you were doing machine learning five years ago. So what are you doing now? And it's not satisfying to come back and say, well, we're doing more effective machine learning. We're doing better machine learning. No, we're doing AI now. Yeah, yeah.

I hear you. Yeah. So I think that's, you know, more or less what I have to say about that. No, that's awesome. That's a really good academic understanding of it, as well as how it applies to the commercial world, which is what most of our listeners come from. So I'm going to sneak in two more Zachary Lipton-related trivia questions, and then I've got one more question for you. So Columbia University, where you got your economics BA, give me the most famous investor graduate from Columbia and the most famous political graduate from Columbia. Those are your questions.

Again, you can email your trivia question answers to rhawkins at compassdatacenters.com, rhawkins at compassdatacenters.com, Columbia's most famous investor graduate and Columbia's most famous political graduate. All right, Zachary, let's go to unethical AI. So as you've given us AI as this, I like you called it more of an umbrella term as I did a little reading, getting ready for us. You hear this term unethical AI and who should stop it. And you, I think, touched a little bit on what's the right use of facial recognition software. Can you give us two or three minutes on how to think about unethical AI, what it means, and what are the questions being asked at the academic level about it? So maybe it's worth stepping back and thinking broadly. So, you know, ethics is not

a property of just, say, an algorithm in the abstract or, you know, questions about ethics or just if you just say, you know, if you spend your entire world, the only thing you think about is that I have data points that come from some 1000 dimensional space and they are separable. And, you know, what is the convergence rate that being able to separate them or identify a hypothesis in some class or something, you're not necessarily addressing a problem that directly maps onto any kind of ethical concern.

However, that's not what almost anybody, I'd say it's the vast, you can't have a vast minority, right? You can only have a infinitesimal kind of subset of the community that's really concerned with, you know, more abstract mathematical questions. The majority of what people are doing is they're training models on real data and hoping that by virtue of training these models, they'll be able to create some kind of product out of it and actually deploy it in the real world. And deployment almost always consists of either making or influencing some kind of decision automatically, right? As we go back to like justice, what is justice? And if you go to like the Stanford Encyclopedia of Philosophy, and you look up like that,

they have this nice long entry on justice. And you see like the sort of central definition is that justice sort of concerns like rendering unto each his due or her due and so it's it's about you know concerns determinations of you know the allocation of benefits and harms in society and some kind of questions about how these relate to what people are what people's rights are. And so, so how does this get back to machine learning? When, when, when do, when, when does machine learning become a, a concern of justice? And, and I think the answer is when it's operationalized to, to, and somehow drive some kind of like actual decision that actually affects people, that actually affects the allocations of benefits and harms,

you know, in society. And so where is that happening? And the answer is, well, all over the place, right? So if you look at all of social media, it's sort of, there's so much crap out there that it's completely unnavigable absent some amount of curation right now. And so what is curation? It's, well, someone uses machine learning to decide what people should see. And so the result

is now someone's ability to be heard is mediated by the choices made by algorithms that are, for the most part, trained in some kind of clumsy and ad hoc way, right? And that's not to say it's an easy problem or that people are being negligent but rather that you know we're basically we're trying to solve a very hard problem that we were not quite equipped for so what we do is we come up with proxies so we say okay i'm going to go train a model to just predict you know is this user likely to click on this item and then i'm going to decide you know make this kind of ad hoc decision that the way i'm going to show you items i'm just going to show you the stuff you're most likely to click.

In so doing, though, you're prioritizing some content over other content. You're amplifying some voices, you're silencing others. Even among people that you might actually follow, you might see nothing that they share. This is just one instance. And I mean, this is actually maybe not the typical pedagogical example that someone would go to. You'd more likely expect someone to see something and talk about the way, well, predictive algorithms are used to provide risk scores. So this is an area that I worked in a bit with my colleague, Alexandra Choldochova, and her student, Ricardo Fogliato. We've done a lot of work looking, and she's done much more, and Ricardo, looking at the use of machine learning algorithms, basically even simple ones, like simple statistical prediction

algorithms, like logistic regression, to train risk prediction models in the context of criminal justice, where basically you have some defendant and you basically collect some number of attributes, you know, how many siblings do they have? What job do they have? What zip code do they have? And, you know, how many prior offenses do they have? What fraction of them were violent? What, you know, et cetera, et cetera, et cetera. You get some number of features.

And they try to predict something like, how likely is this score to a judge as ostensibly some kind of like objective score to say who's a high versus low risk defendant so that they can you know this can sort of inform their judicial decisions right there's some things which are decisions that ai is making autonomously like which content you should see in your news feed where it's just the scale is so large there's no opportunity to have a human like actually interacting in the loop and, you know, making manual curation decisions. There's other decisions like in criminal justice, where, you know, you know, these machine learning tools are being provided as a like, a supplementary piece of information that maybe doesn't directly make a decision, but it influences the decisions that get made,

like in criminal justice. There's other contexts where, you know, like, you know, credit scoring systems or automatic sort of loan approval decisions where they might be getting made automatically based on such a statistical decision, at least for low loan amounts, like consumer loans, you know, or like your loan to help you buy your phone or your laptop or something. And it might be assisting a credit committee who ultimately makes the final determination for a much larger loan that requires some kind of human oversight. But, you know, the high level here is if you look across society at all these ways that machine learning is being used, whether it's in criminal justice and lending, you know, in the propagation of information

through, you know, increasingly the only bottleneck to access it these days, which is increasingly social media, then you suddenly have these technologies being deployed, even if they're framed from a technical standpoint, just as prediction problems, like predict the click from the content, something like that, or predict the click from the content user. That's not exactly what they're doing at the end of the day,

not just making a prediction, they're actually making a decision about who to show what, or they're making a decision about who to lend money to versus not to, or they're making a decision about do you incarcerate someone versus do you not, or they're at least assisting in the making of that decision. And then it becomes a question about, well, who benefits and who is harmed? And so, you know, there's all kinds of ways that you could see how this could start going wrong, right?

Like if it turned out that you're training your recidivism model for criminal justice on some proxy, like, for example, let's say that you were training the model to predict arrest. But we already know a priori that, say, certain demographics of people tend to be more like over-policed. And so even if they committed crimes at equal rates, it would be more likely to be arrested for them, right? Then you'd be in a situation where those people would potentially be disproportionately recommended to be incarcerated, even if they were in equal risk. And so like one area that a bias creeps in here is that

maybe what you mean to predict is who's going to commit an offense, but what you're actually predicting is the data that was available, which is who's likely to be arrested. Right. So there's all sorts of context here. Resume screening is another one. I think a huge number. I think some of us are lucky enough. Like, you know, I think now like one of the nice things about a tenure track faculty position for all the for all the stressfulness of it is that this hope that one day, you know, I think now, like, one of the nice things about a tenure-track faculty position for all the stressfulness of it is that this hope that one day, you know, you'll just never interview for another job again. You know, you'll just sit with your books and read and everything. But for the most part, most of the people in the world have to interview for jobs.

You know, I don't know, my parents' generation, people had jobs. Like, also, I think, you know, many people worked at the same company for life. Now, it's quite a dynamic world and people interview for jobs often. And, you know, every few years and over the last few years, there's been a huge shift towards relying on these sort of automatic prediction based tools for weeding out resumes, for deciding which resumes to pass on to, you know, interview stage versus just, you know, weed out altogether. And so on one hand, you know, this is obviously appealing to would-be hirers because the volume of applications might just be so large that it's very hard to manually look at all of them. But the downside is then the question is, well, on what basis are people's resumes being elevated or deprioritized? And so, you know, you can think of this as like the ethics is not something that is part and parcel to the algorithm itself, but rather that like the status of like being ethically fraught is a property of the scenario,

the real world situation, the real world decision that you're making. Some decisions are constantly like if i eat the train of machine learning model to predict am i going to eat fruit loops or lucky charms in the morning uh nobody cares it doesn't matter what algorithm i use there's there's no ethical import it's because it's just it's just not slightly different important than recidivism rates absolutely yeah yeah here here it's not a concern of of ethics however if i if i'm you know, on the other hand, who goes to jail? Like decisions about like the carceral system are fundamentally serious questions about ethics, whether or not we use machine learning. But because they are, when we deploy machine learning in those environments, we have to understand how, you know, the decisions made by the machine learning sort of line up against those sort of like ethical considerations. And I think that's

the situation that we're in is people get more and more ambitious about using machine learning for, you know, surveillance. You see it being used for face recognition, you know, and there's, you know, difficult questions to face. Like on one hand, I think a lot of people quite reasonably are apprehensive about ever allowing face recognition to be used by law enforcement and concerned about entering the sort of surveillance state. And there's questions about, well, then, who is heavily surveilled, right? And it's, well, if you live, if you own a bunch of property and whatever, maybe you're not likely to be spotted, you know, at any moment. But on the other hand, if you live in like densely populated areas, if you live in public housing, maybe you're more likely to be, your life will be impacted significantly. So it's like there is this shifting of, there is a power dynamic, like that to be considered and how this stuff would be applied.

On the other hand, after the Capitol riot, facial recognition is being used to sort of identify domestic terrorists. And I think, you know, most people think that's a good thing. And so, you know, there are these very difficult questions. And I'm not coming out saying like I've solved all of them or know exactly the right way that every technology should be regulated. But to try to paint some kind of balanced picture of how, you know, these are hard decisions. There's nothing about AI or ML that makes it like magically immune to these ethical decisions because these ethical dimensions are aspects of the problem, not of the algorithm in the abstract. And so when we start making decisions using an algorithm,

we have to think about how the algorithm matches up against various ethical desiderata, just like you would for, you know, any human decision maker in those settings. And the question then becomes, well, you think we have some kind of framework for understanding how humans behave and what motivates humans, what sources of information humans have, you know, like you have some, you think you have some kind of theory of mind for maybe that could help you to understand how judges behave or something like that. Whereas for these, you know, data-driven systems, it's a little bit harder to wrap your head around what the failure modes might look like. Yeah. I love your, your summary to the Zachary of a, it's not the machine or the algorithm that's

inherently ethical or out of itself. It's, it's, it's, it's not the machine or the algorithm that's inherently ethical or unethical in itself. It fits in the context of the larger question, which I think is key to what we're saying here is, hey, what's the problem we're trying to solve? What's the question we're asking? What's the ethics of that? Whether we enhance it with machine learning or AI is not really the key part of the equation. If I capture what I think you were summarizing.

I would say yes and no in that. Like, you know, if you, the AI can complicate the scenario because maybe we have some framework for how to think about a problem based on decision makers that we can relate to. And we don't really know how to parse the ways that AI might go wrong

in those ethically loaded scenarios. But yeah, right. I agree with like the main part that, yeah, it's that, you know, even if the AI complicates things, it is still like in context of a particular situation that is already loaded

with some kind of ethical import. With some ethical factor carried on into the problem without the computer assisted, whether ML or AI. Well, this has been awesome, Zachary. I think we'll have to have you back because you've been so interesting to talk to and have so much good info for us. I'm going to get one more trivia question in that is related to your history. I want to see if folks know what the mascot for UC San Diego is. That is trivia question number four. Again, answer all

four questions. Email us with your correct answers. We'll get you in a drawing for a $500 Amazon gift card. Zachary Lipton, it's been great having you. We are super grateful for all your insight and would love to have you again on another edition of Not Your Father's Data Center. You've been great. Thank you so much. Great. Great to meet you.