Machine Learning as a Software Engineering Enterprise with Charles Isbell - #441
**Full Article Based on the Provided Transcription**
---
### Introduction: A Conversation with Charles Isbell
In this conversation, we catch up with **Charles Isbell**, Dean and John P. Emly Jr. Chair at the College of Computing at Georgia Tech. Known for his work in interactive AI and education, Charles shares insights into his journey, recent developments in AI, and challenges facing the field.
---
### A Glimpse into Charles Isbell's Background
Charles Isbell is a leading figure in artificial intelligence (AI) and machine learning, with a focus on **interactive machine learning**. His research involves building intelligent systems that interact with multiple entities, including humans. This work requires a deep understanding of game theory, human behavior, and engineering.
Beyond his technical contributions, Charles is deeply committed to education, particularly making higher education accessible. He emphasizes the importance of **accessibility in graduate education**, which led him to play a pivotal role in Georgia Tech's **Online Master’s in Computer Science (OMS)** program.
---
### The Online Master’s in Computer Science Program
Georgia Tech's OMS program is a pioneering effort in online education, offering high-quality graduate education at an affordable cost. Since its inception, the program has enrolled nearly 11,000 students, making it larger than the combined graduate student populations of MIT's colleges of science and engineering.
Key features of the program include:
- **Affordability**: The degree costs $6,000, a fraction of on-campus programs.
- **Accessibility**: Admissions are open to all qualified applicants, with acceptance rates significantly higher for online students compared to on-campus peers.
- **Quality**: Students earn the same credentials as on-campus graduates.
Charles highlights the program's success in democratizing education and preparing students for impactful careers in computing.
---
### A Special Recognition: Malcolm Gladwell’s Podcast Feature
In a touching story, Charles reveals that Malcolm Gladwell reached out to him about his father's experience at Georgia Tech. This connection led to an episode of **"Revisionist History,"** where Gladwell explored how Georgia Tech has evolved since his father's time, particularly under the leadership of a Black dean—a testament to progress in diversity and inclusion.
---
### Insights from Charles Isbell’s NurIPS Talk: "You Can't Escape Hyperparameters and Latent Variables"
Charles recently delivered a thought-provoking talk titled **"You Can’t Escape Hyper Parameters And Latent Variables Machine Learning As A Software Engineering Enterprise."** The talk, co-authored with Michael Lipman, was crafted as a narrative to address critical issues in AI.
#### The Core Message: AI as a Software Engineering Endeavor
Charles emphasizes that machine learning (ML) is too significant to be treated as a narrow technical problem. Instead, it must be approached as **software engineering**, involving:
- Understanding the broader implications of systems.
- Collaborating with diverse stakeholders.
- Ethnographically studying real-world impacts.
He draws an analogy between "compiler hackers" and software engineers, highlighting the need for broader thinking when building AI systems.
#### The Role of Diversity in AI
A key theme is the importance of **diversity** in AI development. Charles argues that:
- Diverse perspectives are essential to anticipate unintended consequences.
- Teams must include individuals with varied backgrounds to avoid biases and ensure ethical deployment.
He reflects on challenges faced by leaders like Tim Nickl in Google Research, underscoring the difficulty of being the voice for ethical concerns within organizations.
#### The Need for Humility and Collaboration
Charles stresses that building robust AI systems requires humility. Engineers must:
- Recognize their limitations.
- Engage with domain experts and end-users.
- Foster environments where diverse voices are heard and valued.
---
### Ethical Considerations in AI Education
Georgia Tech is integrating **ethics and responsibility** into its curriculum, particularly at the undergraduate level. Courses on professionalism and ethics are now prerequisites for core design courses. This ensures students consider ethical implications early in their education.
Charles believes this approach will:
- Prepare future engineers to think critically about the societal impact of their work.
- Encourage responsible innovation.
---
### Looking Ahead: The Future of AI
While challenges remain, Charles remains optimistic. He sees promise in:
- **Technical advancements**: Solving complex problems with innovative approaches.
- **Diverse talent**: Future generations will bring fresh perspectives and ethical rigor.
The key priority is ensuring that the next generation of AI professionals is equipped with both technical expertise and a strong sense of responsibility.
---
### Conclusion
Charles Isbell’s insights highlight the dual challenges and opportunities facing AI today. As the field grows, it must balance technical innovation with ethical considerations, diversity, and accessibility. By fostering collaboration, humility, and responsible education, we can ensure AI’s future is both promising and equitable.
---
This article captures the essence of Charles Isbell’s thoughts on AI, education, and ethics, providing a comprehensive overview of his perspectives and contributions to the field.
"WEBVTTKind: captionsLanguage: enall right everyone i am here with my friend charles isbell charles is the dean and john p emlay jr chair at the college of computing at georgia tech charles welcome back to the twiml ai podcast thank you very much i am very happy to be here it is nice to see you again it is great to see you it is so wonderful to have an opportunity to chat with you we do not do it often enough uh we were just reminiscing that you join me here on the show at episode number four uh that was about 440 episodes ago or or even more interviews ago back in september of 2016 when uh i was just getting this thing going and thinking okay whose arm can i twist in this field of artificial intelligence to get on the phone with me get on a call with me for an hour and uh jibber jabber and jibber-jabber we did we had a great time talking about your work in interactive ai and uh i'm looking forward to catching up with you but before we jump into um well another long rambling discussion i imagine i'd love for you to share just a brief refresher on your background and what you've been up to in the past four years or so what have i been up to in the last four years so well so i do ai and machine learning and in particular i care about interactive machine learning as you noted that's where we worry about building intelligence systems that have to interact with dozens hundreds perhaps thousands of other intelligence systems some of whom might be human and there's a lot of sort of math that goes into that there's a lot of sort of theory including game theory there's a lot of effort energy engineering that goes into building those kinds of systems but it turns out at the center of it is trying to figure out the human experience how you take advantage of it how you predict it how you influence it and how you make machines better by leveraging what human beings are good at and there's a lot of interesting stuff there that's what i spend my research time on i spend the balance of my professional time on education i worry about access and how it is that we can educate as many people as we possibly can from as many backgrounds as we possibly can and allow them to be better to actually uh you know manage to get to the places where they want to get to because i think that's the purpose of higher education so that's what i try to balance my time between understanding humans as a machine learning person and understanding humans as a person person and on the education side of things well as well as uh ai you are the no no creator or driving force or you tell us but behind georgia tech's uh online masters in computer science which uh i think did a ton to make graduate education and cs more accessible kind of before it was cool or easy or whatever it is now you know more popular certainly but i still recommend that program because the you know the quality of what you get relative to the the cost is you know remains uh pretty impressive thank you i appreciate that well the goal so we have young masters science computer science we are in our seventh year now i think officially we're about to about to start our eighth in a in a couple of weeks um we went from zero students in that time to almost eleven thousand it is the first online mooc based uh computer graduate computer science program from uh i'll just say elite university at least in the the united states uh the goal is to again provide access to people who want to pursue an education in computer science or in computing more broadly um we admit everyone we believe who can succeed uh so you know on campus we admit maybe 10 of the students uh for the online version we admit something north of 60 of the students and of course i buried the lead the cost of the degree which is exactly the same as the degree that's on campus is about 6 600 as opposed to the 45 46 000 you would pay uh typically if you you came on campus but same requirements to get in same requirements to get out uh same education and the goal is to educate as many people as we can just to put that in sense of scale there are more people in this program than all of the graduate students in the colleges of sciences and engineering at mit combined um as you know georgia tech's the largest engineering college in the country but quite a bit uh the college of engineering which has 45 of the faculty at georgia tech and 41 of all the students uh the college of computing there now with nine percent of the faculty has about 40 of the students and that's almost entirely uh due to the graduate population so the goal is to be big while maintaining quality and to educate as many people as we as we possibly can and to to do what we can to help people to do what they want to do without sacrificing quality um and making certain that people can get out on the other end able to do interesting computing work and i think it succeeded or at least is succeeding nice nice and interestingly enough uh you just learned that a um your work with the the the college of engineering and the online master's program was featured in an episode of malcolm gladwell's podcast revisionist history tell us a little bit about uh that and the connection to your work there so uh malcolm gladwell uh reached out to me what feels like 20 years ago but i guess was just a couple of months and said he wanted to talk to me about something i mean everything in 2020 feels like that many years ago uh asked me if i could uh just chat a little bit with him and i said of course i'm a big fan of his work big fan of the show and he said he would tell me the secret of why he was asking me this a little bit a little bit later uh and um so he asked me a bunch of questions about oms actually and some work that we're trying to do in sub-saharan africa to to connect there with universities and some multinationals to to do some education there and um he told me why he was interested and i i don't want to spoil it for anyone but it has to do with his father and his father's uh interactions at georgia tech and how much georgia tech has had to change uh based on how they interacted with him many many decades ago and how you now have a dean of the college of computing uh who happens to be black uh and it's a i think it's actually quite a touching story and then i found out this morning that he actually released it on revisionist history when i got texts from people yelling at me for not telling them um but it's sure it's already shorter than the amount of time we talked uh so far and uh i actually think uh the story of his father's by touching it so i'd recommend it oh that's awesome that is awesome uh so the main thing that we wanted to jump into here at our time together is your recent invited talk at nurips which is titled you can't escape hyper parameters and latent variables machine learning as a software engineering enterprise uh i'll admit i have not had a chance to watch or listen to it yet i know i know it's been a crazy crazy crazy time um but uh i wanted to take an opportunity to to you know chat with you about it and maybe a good way to to jump in is to have you um kind of share an overview of the talk my my sense from both reading the title and kind of observing some of the conversation that's been happening on twitter as a result of the talk is that you have in a quite skillful way blended a conversation around what it means to do machine learning in the real world as well as uh kind of raise the issue around raise or discuss issues around diversity and inclusion in the field um that's my take uh you know tell us what what you had in mind okay by the way i should tell you um uh this is a little bit of a prelude that one of the i got asked to do this and of course it's an honor to do and i was happy to do it and i thought you know how am i going to make this work giving talks in this environment the virtual environment has been incredibly painful um i think for everyone particularly if you want to be able to interact with the audience um and they said well you're going to record it you can it'll it'll be fine you do what you want to do and i thought okay well if i'm going to record it i might as well try to do something interesting and i've got a few months to do it so if i have a few months and i procrastinate for only three months that'll give me at least a month to put something together so um i got with my good friend michael lipman uh who i taught a couple of classes with and done some other interesting things with and we started brainstorming about what we could do i said this is what i want to talk about how can we make this interesting and we sort of spent months sort of thinking this through had a communications team you know we did all this uh to sort of bring something together and in the end we decided that um the right thing to do was to create a story that would allow a kind of dialogue between people to dive into this really specific and uh i think timely issue around uh bias in machine learning both you know from people who are participating in it but also in the way that is being used and deployed in the real world and so we wanted that to be entertaining because otherwise what's the point of coming together we wanted it to be um it'd actually be informative and we wanted it to be technical but also accessible to people who either don't know about the specific issue or even new to machine learning and which turned out to be both easier and harder than you could have you could have ever imagined so we put this thing together even among in covet and i think it i think it came out pretty well the idea behind it though was the the sort of thrust of it is that uh when i and i say this as a kind of a metaphor by a metaphor and a half that machine learning we tend to focus on the small we tend to worry about the algorithm here's the algorithm that i'm working on i'm gonna try to you know make it solve this particular problem that i think is important i'm not gonna worry about anything else because that's what i do i'm the mathematician i'm the scientist and that's very natural to do it's very natural thing for people to do in all kinds of fields um but machine learning is just too big it's too important ai is too big it's too important computing actually it's just too big and too important for us to pretend that this little thing that we're doing over here not gonna have all these knock-on effects that actually have a real impact on the world uh and so we it's time to grow them up it's time to grow up right that if we think of ourselves as compiler hackers people who are trying to solve a very particular language translation problem we should be thinking of ourselves as software engineers we should be thinking ourselves as programming language nerds it is people who are building entire systems that will involve all kinds of stakeholders and we're trying to describe the things we want to do in a way that is accessible both to us as as experts in machine learning but also to other people who are experts in the domains where these things are going to be applied and finally we have to be ethnographers we have to really understand the fact that this system that we're building will one day get deployed um and we'll have real impact in the real world and it's just it's no we can no longer afford either intellectually or ethically to ignore the fact that we have to build real robust systems so how do you kind of tell that story not just from a you know it's the right thing to do point of view but also that if we take it that way if we treat it that way if we think about it that way then it means that we will actually be able to not just solve more interesting problems but they're technically more interesting things that we could be doing there are different ways we could be thinking about ai and machine learning that'll lead to intellectually stimulating uh possibilities and that we've been ignoring those possibilities in an effort to stay over here in this space and so that's kind of the idea and it's told as a christmas carol with ghosts from past present and future as important is uh and michael litman plays scrooge because of course he does um there's a line in there where michael uh complains that i've cast him a scrooge and i tell him that he's old enough to be scrooge which is kind of a running joke in our lives about how much older he is than i am um turns out he actually is old enough to be scrooge it turns out that he's the same age as scrooge was in the original sort of dixie dixonian stuff i always want to turn it into john dickerson anyway um the uh in the original story uh because it turns out that in your mid if you lived to your mid-50s you lit a hard life uh way back when and uh looked a lot older so actually scrooge apparently was in his 50s totally appropriate i think so i think so michael's lived a hard life and i think so uh in there you kind of start with this analogy of kind of compiler hackers to software engineers i'll unpack that a little bit and what that all means yeah so by the way when i say compiler hackers i know there's lots of compiler hackers now who are mad at me but i need something very specific about that right it's an important thing that that we do and i'm just going to pretend i can just describe it as a machine translation right we're worried about solving a particular problem it's relatively well specified there's lots of really cool interesting things going on there uh but it's it's a kind of world where you have defined what it is and you know what the output needs to be and you're worried about all kinds of interesting optimization problems it's not that it's not interesting but it's narrow in a good sense but it's narrow and sort of it's focused once you start thinking about systems you have to start bringing in other things right people will talk about system you know i'm assis i'm a software engineer oh so then you do coding no a software engineer has to actually understand what specifications are they have to understand what requirements are they have to understand where things are coming from in our case that's the data the problem and they have to understand how the systems are going to be deployed ultimately that's the models that we produce if you want to think about it in terms of machine learning or the larger systems we produce if you want to think about it in terms of ai and if you don't gather that entire pipeline if you don't think through the entire system of which you might want to think of yourself as a small part then you end up building systems that don't work or worse they work really well it's something you didn't intend for them to work at and they go out in the world and they cause real harm and that's intellectually dishonest and lazy and uh it's dangerous and in a world where these things are being distributed everywhere and they're being used by everyone we can just no longer afford to do that and so when i talk about the difference between software engineers and compiler hackers what i really mean is the difference between working on this interesting problem versus having to deal with what happens when you have to be a part of a much larger system including large parts of the system that you do not understand you're not an expert in but nonetheless have to have in your head as you're trying to build something you have to work with others it's not possible to do it by yourself you have to bring in all these other people you know before you show up and after after you're done and that's what we have to do and that's really the kind of conceit that's really the sort of the idea here and i think that we avoid it i think and there are good reasons for this but i think that we avoid that reality because it's a lot more work um it feels like doing things we don't i'm not trained to do you don't necessarily want to do but i actually claim that it's uh it's better it's actually better and it's actually more interesting it's just a more interesting set of problems that have the nice side effect of being better for humans as well yeah i experienced a little bit of dissonance with your analogy living pr predominantly in kind of the the tech world you know which is dominated by software engineering and software engineers um but it's still often characterized by uh you know monoculture you know move fast break things a disregard for the broader you know systemic implications of what they're doing and in some ways you're still holding you're at least using the same term as an ideal that we need to move towards in the mlai community does it apply more maybe in an academic sense no no no i mean i take your point i think you're absolutely right it's very easy to think of yourself as a software engineer and still be unnecessarily narrow uh in the way you think about the implications of what you're doing and it's very narrow again to do that i'd certainly have been guilty of it and i suspect that i'll be guilty of it over sometime in the next couple of months um but you know it's an ideal right it's aspirational but actually more than being a woman being aspirational if you take it seriously right there's two things you should you can think about one is i'm building a system and if i build it wrong bad things happen so you know there's lots of examples of this uh there's a book set phasers on stun i think it's called you know you build a system that is going to deliver radiation for x-rays and it turns out the ui is terribly built and you literally kill people right once that happens you have to start building systems if you have to start you have to build a methodology uh and you have to take care to think through how the system's really going to be used understand the user you have to really build out and out and now software engineering is about that there's lots of things that are about lots of systems thinking that about that it's hardly um limited to computing but that's a part of what i mean and that's one thing that you get out of it but there's a second thing that you get out of it and this is the sort of thing i think you're talking about that not everybody does which is it's not just a matter of here's the user i imagine and i'm going to make certain that that user is going to be okay it will be a nurse in a specific situation and no more the problem you take that idea and you apply to what we're doing in machine learning and ai and computing in general it's not a nurse in front of a machine providing um a service as a part of doing an x-ray it is a random person who's one of eight or nine billion people who have access to this technology and will use it so the idea still applies but now you have to worry about a much broader set of users you don't get to you know say well these people don't don't get involved because the system is going to be out there and it's going to be used for all kinds of purposes you haven't necessarily imagined what that leads to and i think this is you know sort of the second half of the discussion for me is it means that you have to involve as many people as possible with as many perspectives as possible and experiences as possible in order to make certain that you're actually going to build these systems that will uh in fact be useful in the world in the way that they're actually going to be used right so that means diversity right that means bringing in people you wouldn't necessarily bring in or wouldn't necessarily want to think about as your users uh bringing it having them as testers and more importantly having them as people who are machine learning experts i think an easy mistake to make and an easy lesson to draw that i think is the wrong one is well you know it might get used by these people who don't look like me or come from a different background so i guess i should visit them and figure out what they what they want and then it'll be okay no you have to think of people not just as domain experts outside of what you do you have to make certain that people from these kinds of backgrounds are also the experts in the machine learning the ai the computing the software engineering and all the other parts of what we do in the first place to help to design the problem someone has to be in the room to say maybe it's not a good idea to have the robot kill people someone has to be in the room to say that's not a you you say you're talking about crime but you're not talking about crime you're talking about arrests those aren't the same thing and let me tell you what my experiences have been like you need people in the room who are technologically savvy understand as much as you do about whatever problem is you're trying to solve from a kind of ai machine learning computing point of view but also had a set of experiences that will help you to think more broadly about the applicability of what it is you're gonna do um and i think that that's the part that one can skip around but without that we don't actually solve the problem we just give ourselves one more step of illusion that we are yeah yeah it's uh i think a really interesting time to make that point um about uh you know needing the you know needing that alternate perspective in the room i think you know the time we're talking about this you know recently tim nick gabriel was uh what resigned i forget the word that they made up to describe it designated resignated from uh from google research uh and i think we're all getting a sense of both how hard it is to be that person in the room um and also how hard it is to actually have that person in the room and allow them to fulfill their duty of calling out the you know potential issues with what you're doing um take on that thoughts on that that's exactly the right point so one of the things that the literature is very clear on is that um you get better outcomes when you have a diverse set of people in the roof and may not like that or you may think that's great but it doesn't matter the data or the data you get better outcomes for lots of reasons but the other thing that the literature tells us that people don't tend to talk about is that those groups are less happy right because maybe it's because they're being challenged maybe it's because it's uncomfortable maybe it's because you know there's more fighting i i did i won't i won't bore you with the details but i did this uh exercise several years ago with 60 something other people uh and they had had us divided up into these little groups um which they had done on purpose we didn't know but they over like a whole year uh and we were working on these projects together and there were some people in my group i liked and there's people in the group who just got on my nerves in a way i just can't even describe you i mean i like them and respected them but man you really have to sit here for 15 more minutes and talk about you know your daughter or whatever i just i don't care i want i want to do this problem and um you know they then brought us together about halfway through this this year and they divided up as again and they divided us up based on personality traits they had decided that we had and it turned out i got put in this in this particular group of people and it was the most fun i had had in the six months that we've been doing this i mean we were completing each other's sentences i mean we were joking everyone was laughing they actually said you know we've been recording you and you'll notice this is the loudest any of you have been in the six months that we've been doing this and all the other times you work together you're much more quiet and it's because they had put us with people who thought the same way that we did have the same experience as we did it was fantastic it was great but the different groups they broke us up in had completely different takes on the problem that had been put before us we saw the world completely differently and each of us was worse off for having not had the perspectives of other people who tended to think to think think differently and i got a lot from that experience and one of the things i got from that experience is man it's easy to surround yourself with people who are like you and it's so much more fun and man it is irritating to surround yourself with people who think very differently from you but you have to do the latter in order to succeed and so what we don't do is we don't do the hard work of saying we're going to bring in these diverse voices and we're going to deal with the fact that it's going to be somewhat less comfortable and we need to create structures and ways in which we can make people both more productive and happier as a part of that and and i think that you know it's worth it it's clearly worth it i mean i'm not going to say anything specifically about what's going on inside google i'm not in there i don't know what's happening i've done enough hr stuff and manage enough people to know that there's always 5 000 sides to every story even if it only involves two people but you know it is certainly the case that it is easy to create an environment where everyone is unhappy unintentionally just by bringing people in and not doing the next step of thinking through carefully how you make that work it does raise some questions around how to kind of operationalize or put into practice some of the the things that you're calling for given the level of discomfort that's created by having the you know the diverse voices in the room and the you know folks that are you know willing to say hey maybe we shouldn't be doing this or just don't do this um what what also needs to be in place so that the organization can take advantage of those voices i mean you know it's hard if i knew i knew the answer to that i'd probably be doing something different but i think there's some stuff you can do right i mean you set expectations i mean you you make spaces i hate to use the phrase safe but you make spaces safe uh you tell people that you know it's okay for and by the way this goes from the top down right as a leader in the room as a manager or a vp or dean or chair or professor whatever you have to say it is okay for us to have these um difficult discussions and it's gonna be okay we're all gonna be friends afterwards and we're gonna move forward because we're united by something there's some goal we're trying to accomplish and this messiness here will get us to a better place and none of it's personal um and if you can get people to buy into that uh and you're willing to take hits yourself as someone who might be um uh say leading the group then people will do it um and they typically are okay you have to find other ways to keep people talking though if you allow people to get upset and to go off in their separate ways and sort of stew in it things eventually get bad but if you have meetings if you have regular things where people get to say out loud what's bothering them um and i don't mean this as a kind of let's all hold hands and share our feelings since just uh hear the things i'm trying to do oh i didn't realize you were trying to do something similar how might we solve this you just kind of keep the conversation constantly going i found that in general those things those things work better but in the end it's people and people are people and people are going to do what people do and you just have to you just have to accept it but if you can create expectations where people know this is a place where you can speak up and by the way the price of you speaking up is that other people will speak up back but in the end we're all going to still get to the same place you can you can usually make that work at least yeah yeah so return returning to your nurip's talk you you said you started out with kind of an exploration of the past what was that uh it was about software engineering it was about theory uh the things that were done there's a really nice example in there michael kearns uh many of you will know who michael kearns is uh he wrote the ethical algorithm um co-wrote it he uh and is a one of the founders of uh theoretical machine learning or at least one of the people who did it in the early days um and a great guy and a very good basketball player and squash player he's kind of annoyingly good at everything anyway uh he's a real nice guy um he talked about how cyber security was like this game you know just terrible it's like this you know i come up with something someone comes with something else and you just kind of never you never get anywhere until people start defining uh their terms very carefully thinking well what is what is it we're actually trying to accomplish what is it we're actually trying to to get right what help us what how do you define your terms and once you put something on um sort of stable theoretical and algorithmic grounds you can actually make progress right and so there's a lesson there combined with what we know about software engineering and and how you can make mistakes if you aren't careful so you just think about what we've learned from the past you realize at the end of the day clear definitions methodology and thinking beyond the problem that happens to be immediately in front of you are all necessary in order to make significant progress that actually impacts the world or at least it seems that way to me and that's what the past was mostly about it was also about laying out the arguments against caring about this the most common one is uh it's not about the algorithm it's it's the algorithm isn't biased the algorithm doesn't have problems it's just the data and if we had the right data everything would be fine and so therefore i can pay attention to just the algorithm but it turns out that doesn't make any sense first off the it is your problem if your data has problems and you need to deal with that secondly um if you know or believe or worried that the data you get is not representative of the problem you actually want to solve the how things are going to be deployed later then you need to build algorithms that will understand that and will react accordingly and if you don't do that then you aren't actually solving the problem you claim to be solving you're solving a made-up proxy problem that won't actually be as useful it might get you a paper and nerves but it won't have the impact that you purport to want to have and i think that's actually a sort of key and important thing to to think through but also you know algorithms have bias by definition i mean i think it's a different kind of bias from what people mean uh when they're talking just globally but um they do and algorithms have hyper parameters and either those hyper parameters can affect what are what is learned or they can't they can't they're not they're not useful they don't mean anything so if they can then it actually does matter how you set the knobs and you tweak things for your algorithm to get the output that you would like to have and so you can't avoid that even though we do look i i had someone call me up once um and asked me about some work i had done in the late 1990s early 2000s in reinforcement learning and they said so what value of lambda did you use uh or gamma did you use and i said oh um i think it was 0.7 why did you use 0.7 uh because that was the default in the code that i had right it wasn't and it worked and because it worked i didn't have to care about finding the right um gamma right uh we all do that right we we have all these hyper parameters they're supposed to mean something uh we kind of do whatever it is that works and that doesn't necessarily mean it's going to work out in the world uh and that's really easy easy set of traps to fall into and it's not a very software engineering view of the world but it is about you know overfitting to the problem in front of us and that's supposed to be something we don't believe in is aip yeah yeah and what's your take on kind of where we are and what where the priorities need to be in terms of defining the the common terms and methodologies that you that you referenced you know there have been some initial efforts around you know things like model cards and uh data set uh data set spec sheets that kind of thing um seems like there's still a long way to go there oh we're we're very far away but i mean that's that makes it exciting right i mean i would just love to have a definition of fair that makes sense um talk to me about this and it is just it is amazing to me how you can come up with multiple definitions of fair all of which seem fair uh but lead to completely different results and just understand outcomes and just understanding that is in and of itself interesting i think you can spend you know there's there's certainly multiple phd dissertations just in the question of defining what does it mean to uh have accomplished some notion of fairness or balance or pick whatever words you want to do never mind the sort of practice of how you would you would get this kind of implemented in the real world so i think we are go ahead not just english definitions but mathematically sound definitions of fairness that uh are very different we haven't even touched on culturally grounded notions of fairness right right we're going to come up with something that makes sense in i don't know the united states but does it make sense in some foreign land like canada you do very weird things like they use this thing called the metric system i don't know it's very strange people out there in the world and you know you think you're like them you think you're or you think you're different from them and just figuring out how you're going to make these kinds of things work and truly generalize is actually very very difficult and very difficult for an individual engineer scientist computationalist whatever you want to call yourself to even imagine and so to me and this you know you know me you know me well and some people will joke that this doesn't sound right coming to me from me but i mean it um you have to have humility about what you know and what what ability you have to actually predict how something's going to be going to be used in the world and but just coming up with those sort of technical and mathematical notions of what it is we're actually trying to do i mean that's that's exciting stuff and i i know we've got decades of work to do in that space uh and we haven't even got to the part of how you're going to bring the people in to even point out to you that the definitions you're coming up with don't make any sense uh in their environment or in their world or in their experience yeah yeah yeah are there analogies that you've seen in other areas of engineering or software engineering for how these uh not necessarily these kinds of challenges but similar types of challenges have been addressed yeah i mean so my you know my observation and i'm hardly a you know one of the problems with being an expert in one thing is that you realize how much you're not an expert in something else so this is not my area of expertise right but my observation of the way engineering and science kind of develop over time is they're usually forced on you by some major disaster and you overreact to the disaster um and then often good things come out of that the pendulum swings and it does whatever does whatever it does but you know the examples that i always kind of think of in the back of my head are things that happen in aerospace right like you you can't get a phd in aerospace without i mean all your you think about safety constant right it's it's built into what you do and that's a good thing it turns out i was at this um oh i don't know it was a convening of people about ai at the white house or something and uh they brought in these people from these different backgrounds there were people from um health and medicine specifically uh there were people from you know boeing you know the the airplane people uh and there were people from various forms of academia they're all these uh accounting they're all these people from these different areas and you've got these people in a room and they're trying they're talking about ai and it takes you 15 minutes to realize that they're all using the same words and they're using them radically differently and it's because they're context of what's important like you you cannot be in the business of building planes without thinking about safety all the time that's all you that's the single most important thing probably at least that's what it looks mean but in medicine you're worried about privacy you're worried about you're also worried about safety but privacy matters there's much less there's much more wiggle room about what does it mean to be successful or not you know chronic versus acute i mean all these things are very different and they're just the problems they're caring about and the way they even think about the world are so so very radically different so when i look at other other fields and other experiences what i get out of it is that they're also very different in the things that matter to them um but they're mostly driven by disasters things that happen that force them to say you know we will do no harm or force them to focus on um privacy or focus on uh safety or focus on this and i would i just think the world is better if you think about it before the disaster happens because if disaster happens then you have very little control over what's gonna drive your field for the next five or ten years because other people are going to come in and tell you so the next part of your talk was the present tell us about the present well that is the problem one of the things about the talk is that you realize halfway through that the past the present and the future are all kind of kind of uh sort of pushing together the present is where we are right the present is the fights about whether we should be whether we should care about this at all right it's you know there's a lot of good work that's out there as you said you know um a lot of proposals about how we think about things like bias for example but we're still at the point we're arguing about whether that's even a problem not much less whether it's a problem worth considering right so you know you will recall it was just a few months ago that we had this big blow up on twitter uh around pulse and style gan uh that would turn people um of color into people of somewhat less color uh there's a picture of me floating around other that just deeply it started with the bomb and it ended with me that's how i kind of think about it you know i just figure obama charles we're roughly on the same level of importance when it comes to this at least as far as the algorithm is concerned because the algorithm turned us both into um very different looking white people uh who looked mildly related to some distant cousin of ours uh in my case it gave me blonde hair it was all very it's all very disturbing um so the present that we were in is really an argument about whether we should be worrying about this at all and and i think it's i'm personally looking forward to getting past the present uh into just agreeing that these are interesting research problems that we should be we should be focusing on so that's the present for you the present to me should be short because it's where we're living right now and we need to get past it and move on into the future and the future which is also the present yes and it's the past it turns out and the past it turns out and so the did you did you provide any uh any crystal ball uh insights into the the future or more directionally the things that we need to be thinking about the only so there are two there are two parts in the in the future and and one is a bit of a turn so there's kind of the technical question of the future uh and those are just sort of laid out um here are the kinds of places that we could be spending our time thinking about here the interesting problems we could be doing some of them i think are kind of obvious some are some are less so but for me the real future is long term and it's less technical in other words the point isn't here's what you should be focusing on algorithmically or here's the mythology that you should be working on that i'm going to push on you because i think it's the right thing the future is really creating the set of people who are going to be prepared to lead the conversation right so we need to be looking at the people who are 10 years old right now because they're going to be the ones who are going to solve this problem in 20 years so that's about education that's about getting people in to realizing that this is important in their lives whether they're going to do ai and machine learning or they're going to do something else but just to be facile and thinking about things like this so that they can make reasonable decisions so to me the picture is both the technical one and that's interesting but it'll come on its own but the one that won't come on its own is making certain that more people are engaged in the conversation that they have the technical background they understand what computing is they aren't afraid of taking derivative because someone told them they should be uh and that they were part of that conversation and that they're in fact in the room they're in the room when the conversation happens and to me that's actually the hard part of the future you get enough people in the room then the technical stuff will come because there's smart people out there trying to solve difficult problems that'll just happen another necessity will will do the right thing but if you don't have the right people in the room you're not even going to ask the right questions to to look at the interesting problems much less find the right solutions that's the thing i worry about i worry about is whether we're going to get the right people in the room 5 10 20 years from now yeah yeah well taking a step back from the the 10 year olds are there things that you're doing or looking to do with the omcs program to you know as an example kind of incorporate you know coursework or classes around ethics in particular or um you know humanities you know more broadly social sciences more broadly is that uh you know is that already in place or is it a direction you're looking at do you think that's important it is in place uh and it is happening so we have been developing courses at the graduate level for this but i think even more importantly we've been doing it at the undergraduate level so georgia tech for decades like over 20 something years uh has always has required what we call an ethics course as a part of the bachelor's of science of computer science uh that ethics it's called an ethics course it's actually more of a professionalism course which i actually think is somewhat more important um in fact i don't tend to talk about ethics i talk about responsibility one of my goals is for georgia tech to be number one in responsible computing and i think responsible computing includes ethics but it includes professionalism it just includes thinking about the consequences of what you do and taking ownership of those of those consequences uh so that you can build systems that are more robust and ultimately ultimately better so why am i bringing that up because this notion of responsibility which has been built in again into coursework we've had at the undergraduate level for for decades now we have made some significant revisions to them over the years and are in the middle of doing something that i actually think is is quite important so i'll take a second to tell you about it so the first thing is we have this class is actually the first class i ever taught at georgia tech was around this notion of professionalism and ethics i just taught it once but it was a great introduction to teaching and computing for me we have expanded that so there's a focus on computing there's a focus on robotics there's a focus on privacy there's a focus on data there's all these different options you can do as kind of contexts because that's what we do we do contextualize learning the sort of context you could use to think about uh responsibility ethics and professionalism so this year we updated our curriculum because what was happening is people would take that course and to be the last course they take and what's the point of waiting until the very last thing you do to start thinking about ethics one that's a very strong signal about how this isn't actually that important um and you know you have vague memories of being a senior uh you know what this is like you're buying somewhere else you're doing other things what we did is we decided it's not only going to be a required course it is a prerequisite for our junior design sequence so we have a year-long junior design course you know one semester is all about um design and the other classes about other semesters about implementation it's a group project you learn all kinds of things i won't go into details we've done all kinds of things about how to build systems starting next year the requirement for the ethics course one of the ethics professionalism responsibility courses will be to be a prerequisite for that so before you get through junior design you're gonna have to have already thought hard about this not your first semester because you don't know enough to have context to realize what what's possible but not to your last semester when you know you could have you didn't have you wish you had had that context in order to think about the things you're doing but halfway through we've got enough background we're now going to make you think about these things you're going to be using it when you do design and sort of start asking these right questions about implications and responsibilities of the systems you're building and potentially deploying and then using that before you take machine learning before you take intro to ai before you take uh advanced cyber security all of these things that that you're going to be doing um uh you will be thinking about uh very early on in the career the idea is to integrate this in and i think that you know without integrating into the curriculum you don't you neither send the signal that it's important but you also don't give people the opportunity to put it in practice in an academic environment where you can actually explore so that's what we're spending all of our energy on it's about responsibility um respect to computing broadly speaking and do you find that pushing that forward changes the the conversations that are happening in the later you know coursework and project work so that some of these issues are recurring in uh you know specialty courses as opposed to just in the responsibility course so that's the idea i mean you know ask me in five or ten years don't have enough data to be sure but even who have done this stuff earlier you definitely see it happening and and the students claim at least this is what they want so i think um they'll be engaged uh in having those those conversations um this is the current generation i think thinks about these things very differently from even the you know half generation before uh sort of thinks about their things about their responsibility but imagine i mean our entire curriculum is designed around this idea that if we can give you a proper context by the time you get to be a junior or senior you are situated in such a way that you can ask interesting questions and do interesting work well responsibility professionalism has to be a part of that or otherwise you're saying it's not an important context for doing interesting things so you're taking a machine learning course and you're dealing with data and supervised learning or reinforcing or whatever it is you're doing in your machine learning course and you haven't yet thought about the implications of data and privacy and how things can get how things get used well either i have to teach you that in machine learning or you're just not going to think about it until it's too late so my expectation is that if i know that by the time you take the intro ai course for me that you have had this experience then i can integrate it into the assignments i can integrate it into the discussions because i don't have to teach it to you um for two or three weeks before we can get into the other stuff that i'm doing and so it'll actually have the side effect i hope of allowing faculty to think about how what they're teaching what they're doing their passion is impacted by these notions of responsibility without having the pressure of being a philosopher who understands ethics and trying to teach it from scratch to a bunch of students and even if the faculty don't particularly want to do it i'm hoping that by the way that we teach this the students will bring it up themselves and the conversation will naturally lead to a better curricular and academic outcome yes i'm very optimistic that this is the right thing uh and that ultimately this is what everyone will be doing and it'll just be kind of built in and 20 years from now people won't even be able to imagine a time when we weren't doing that awesome awesome well you know throughout your talk you've raised a ton of issues asked a lot of questions for which we don't have good answers um have you did you kind of wrap things up with any you know we should do prescriptive uh next step for folks or is that an exercise left to the observer well it's an exercise left to the observer but i will tell you the thing that i did leave with him and i think this is this is true um certainly it's true of me in the end i'm optimistic i really think that you know we're living in difficult times um but then when have we not been living in difficult times uh for the field for machine learning and ai and computing actually generally speaking uh we are living in difficult times we are having to mature and wrestle with the fact that the things that we're doing are actually impacting the world and it can go very wrong but the point is not that there's a horrible future in front of us that we have to avoid it's that there's a wonderful future in front of us that we have to embrace and walk down that path as quickly as possible so i come out of the whole experience and not just from a narrative point of view but i genuinely believe this thinking that we're going to end up in a good place there's going to be some bumps and there's going to be some mistakes but we're going to be okay and i think that we're going to make the right decisions at least in trying to think about these problems whether or not we make the right decisions every day about whether to deploy this or whether to set lambda or gamma to 0.7 or 0.4 we're going to do the right thing at least i'm hopeful awesome awesome well charles it was great catching up with you i hope we did not steal all of your thunder uh you should definitely check out charles's talk i certainly will um but you know once again it is always a pleasure to catch up with you yeah it's fun let's uh let's not wait another 440 episodes to do this again for sure for sure all right awesome thank you thank you\n"