How to land an ML job: Advice from engineers at Meta, Google Brain, and SAP
Check out this video, summary and transcript of a discussion between co:rise co-founder Jake Samuelson and three outstanding ML engineers — Kaushik Rangadurai, Shalvi Mahajan, and Frank Chen — to hear their advice on landing a job in machine learning.
co:rise co-founder Jake Samuelson recently sat down with three outstanding ML engineers — Kaushik Rangadurai, Shalvi Mahajan, and Frank Chen — to get their advice on landing a job in machine learning. Kaushik is a technical leader at Meta, and has over 10 years of experience building AI-driven products at companies like LinkedIn and Google. Shalvi is an AI scientist at SAP, and has experience as a data scientist, a software engineer, and project manager. Frank is a founding engineer at co:rise and started his career at Coursera, where he was the first engineering hire and built much of the platform’s original core infrastructure.
The following excerpts from Jake’s conversation with Kaushik, Shalvi, and Frank have been edited and condensed for clarity.
You can watch the complete recording here.
Jake: Let’s start with a really practical question. Kaushik, you’ve been a hiring manager at some big companies. You get a lot of resumes. What are you looking for? What advice do you have for someone who’s working on their resume and thinking about how to position themselves?
Kaushik: In terms of skills, I’m looking for a practical knowledge of applying ML to build products. That’s something I think you can’t get from books — you have to have some hands-on experience. I’m not necessarily looking for someone to have experience with specific tools or techniques, because those things are constantly changing. It’s more that I want to know about the approach they took. Why did they use the tools they did, and what did they do when things got tricky or didn’t work the first time?
Don’t get me wrong, I think having a good theoretical foundation is definitely necessary. But I would say you should spend as much time as you can solving real problems. That’s how you learn which techniques work best for which use cases, and it will help you get a better understanding of the theoretical side, too.
Jake: That’s great. Any other advice on preparing for new roles, or preparing for interviews?
Kaushik: In terms of preparing for interviews, other than brushing up on the fundamentals, my advice would be to brainstorm a couple of problems that are relevant to the company you're interviewing with and do some background research on the common techniques to solve those problems. Like if you’re interviewing at YouTube, I would try to go in with some thoughts on their recommendation systems. What are the issues you're seeing in YouTube recommendations right now? Where would you start if you were tasked with making improvements in those areas?
Jake: Frank, let’s jump over to you. You've worked at big companies like Google, and also some smaller startups like Coursera and now co:rise. What advice do you have for people who are weighing those two types of opportunities and trying to figure out what’s going to be the best fit for them and their career?
Frank: I think everyone should experience both at some point. They’re just really different. At a big company, you’re probably going to be working on one tiny piece of a huge, complex system. Iteration cycles are going to be pretty long. You might encounter more bureaucracy, and there might be a long approval process to actually get your things to launch. But there’s also going to be a lot of impact from your work just because of the sheer scale of the products that these companies put out.
On the other hand, in a startup, you get to build a lot more things a lot faster. Everyone’s just trying to get to product-market fit and get to cash-flow positive. It's a lot more hectic, but it can be a lot of fun, too. To anyone who’s still in school, I would recommend finding a couple of internships, one at a big company and one at a startup. That will help you figure out which is a better fit for your personality and the lifestyle you want to have.
Jake: We have a couple of questions from the audience. One is on the idea of starter projects. What advice do you have for people to get hands-on experience if they’re not currently in an ML role? I know that can be hard if you don’t have access to data or you’re just not sure which direction to go in to make sure you’re working on something with real-world relevance.
Shalvi: Kaggle is a great resource. They have real datasets and challenges based on real-world problems. Their GitHub repository has a lot of documentation and explanation of the way they suggest approaching those problems.
Kaushik: I’d also recommend looking at some existing papers that have been published in this field, and then digging into the datasets that were used to write those papers. The advantage of doing that is you'll have a good baseline for the metrics the paper is evaluating.
Frank: One thing I’d add is that if you want to be in machine learning, and you have more of a software engineering background but not much data science or machine learning experience, you can look for adjacent roles in software engineering. There’s a lot of software engineering involved in getting machine learning systems into production, and that’s just about being a good software engineer rather than knowing a lot of math or having published papers or anything like that. Getting into one of those adjacent roles is a really good way to learn without, say, having to go back to school.
Jake:. I'm going to take another audience question that feels related to what Frank just said. The question is — is machine learning a young person's game? When you look around, you do see a lot more young people in this field. Can any of you share an example of someone who’s transitioned to ML later in their career?
Frank: I don’t think it’s a young person’s game. You do have to keep up with new developments, because the whole field is changing so fast, but that’s going to be true regardless of your age or where you are in your career. You just have to budget time for it. Almost all or the researchers I know spend a lot of time reading and reimplementing papers. I think that just becomes part of your job description.
Jake: We have a few people who are asking about kind of a day in the life. What do ML engineers actually do? Maybe you all could give some examples of what you spend all your time doing day to day.
Shalvi: I can give a couple of examples from my work. I’m mostly focused on validating machine learning use cases. So it starts with looking at a problem and deciding, do we need machine learning to solve this, or not? If the answer’s yes, we do some brainstorming on how to approach the problem and build a model for it. Once the model is finalized, we do some business value assessment and evaluate some performance metrics, and then we decide whether we’re actually going to put the model into production. It’s actually a lot of brainstorming sessions and meetings in addition to the research and development pieces.
Jake: ML is getting to be a really diverse field with lots sub-specialties. Which areas of do you think are maybe a little easier to get started with? Obviously you should follow your interests, but practically speaking, are there areas that are more straightforward or easier to get into for someone who’s totally new to this?
Kaushik: It’s a good question. One thing to be aware of is that right now, there’s a lot of convergence across different fields within ML. The techniques you’re going to learn and use in one field are often going to be applicable to a lot of other areas. So from a skills or knowledge perspective, you don’t have to pick a direction right away.
That said, I do think search is one of the easier fields to get started in. The basics of search are pretty straightforward. You start by working with short queries that are easily understandable. You can see the results, debug things, etc.
Jake: That’s good to know. And I’ll put in a plug here for the two-course search track on co:rise. If you are looking to get started in search that’s a great place to go.
One last question. As you all have said, so much is happening in machine learning right now. How do you keep up with trends and new developments? What are some of the ways you manage your time or remind yourself to pick your head up and look around, so to speak, especially when you’re dealing with the day-to-day demands of a full-time job?
Shalvi: I really like subscribing to certain news alerts, especially for big companies like Google or Meta. That makes it easy to read about what’s going on in technology. I’m also connected with SAP blogs, and working with co:rise has been great because our courses really are focused on recent developments and tools. And then it helps that a lot of this is in my job description. I need to read papers and stay focused on new technology to do my job.
Jake: Frank, how about you? You always have a couple of different monitors up, I feel like you’re tracking what’s out there all the time.
Frank: A lot of it is following Twitter and things like that. Finding researchers who you admire and who post or repost interesting things. I couldn't possibly go into their archives and read all of their papers, but I find that just skimming what other people are saying can be really helpful. That way I have an awareness of what’s out there, and when a certain type of problem comes up in my work, I’ve already thought a little bit about how that new research or those new tools might help me solve it.
Jake Samuelson: Awesome. Recording's in progress. Got my hair did. Well, thank you everyone for joining us. Welcome. We're so excited to have this panel together to talk. I guess jump-starting your ML careers, I am not sure, so hopefully, no interrupts on the call now. But I'll do my best to host this panel. I'm just checking, yeah, no one dropped a call. But the stars of the show are Shalvi, Kaushik, and Frank but I'll introduce them in a second but just quickly on Co:rise, we are an education platform that is aiming to transform the way professionals build technical high demand skills and we do that by partnering with world-class instructors so people really at the top of their fields, like this panel and we do that as well through peer and social learning. The format of the courses for those of you that are not currently in our platform is a mix of live instructor sessions, real-world projects so most of the time you're spending time building hands-on keyboard, not watching videos and taking quizzes, but actually building projects that would mirror what you would do on the job and then we have a lot of fireside chats and discussions with operators who are experts in their fields, bringing in different perspectives like this chat today.
PARAGRAPH: The goal really is that everything is extremely practical so people either running back to their desk to apply what they've learned or maybe you're taking your projects and using them to talk through an interview as you think about transitioning careers. Very excited to introduce Co:rise and now really onto the show. I'm going to go in order of who I'm seeing but Kaushik, I will not give your extended bios but your short bios and you can tell us a little bit more about yourself. But Kaushik is a technical leader and content understanding team at Facebook. He holds several patents and publish papers and search recommendation, has over 10 years experience building AI driven products at LinkedIn, Google or Microsoft and as an early engineer at Passage AI. So thanks for joining us, Kaushik. You can wave your hand. Cool. Awesome.
Kaushik Rangadurai: Thanks for having me.
Jake Samuelson: Yeah, awesome. Shalvi is an AI scientist at SAP. She has experience being a data scientist, a software engineer, project manager and her focus is really on deep learning, machine learning, NLP and data mining techniques. Wave Shalvi.
Shalvi Mahajan: Hi, thanks for having me here.
Jake Samuelson: Of course. Frank, last but not least, so Frank started his career as the first engineer at Coursera. So Frank, was a fun fact, was a sophomore in college when millions of people were using his Mooc platform and he was running between making sure the platform was up and finishing his exams multiple times a day. Then he continued on to build Coursera's really core infrastructure before moving onto
Google Brain and now is a founding engineer at Co:rise. So thanks for joining us Frank.
Frank Chen: Yeah. Hi everyone.
Jake Samuelson: Awesome. Cool so let's jump right in. What I wanted to cover first and then maybe I'll go in order, maybe I'll start with Shalvi. Love to hear each of your journeys into building data and ML systems. Maybe talk about how you got here to be doing this in your current roles. Shalvi, I'll start with you.
Shalvi Mahajan: Thanks Jake for handing that over. First of all, I did my bachelor's in computer science and engineering and I started back in 2013 doing that. The first time I got that experience of working with data and data analysis was when I did my internship. I did my internship in [inaudible] in south France. And there I had the task or the project was basically based on detecting telephone and fraud and had to do data analysis for the same. That was when I realized that this is really fun, very interesting topic, at least that interests me and I thought that I could definitely do something more about it. But then I got into Samsung. I got my campus placement and I started working at Samsung as a software engineer so that means I was more focused on software development aspects, coding, and stuff but my data topic was nowhere there. I wanted to explore it more then I came for masters at Tum Munich just after one-and-half year of working at Samsung.
Shalvi Mahajan: There in my masters, I started data science. Here I was only focusing on the concept of data science because I knew I was happy with my internship, but I also wanted to make sure that I have the most important skills and concepts I gathered from my masters. I think the most fascinating stuff for me was that data science is not computer science. Computer science is just a tool helping for machine learning techniques and so on. But machine learning is mostly about mathematics. That was when I realized I need to make my basic super strong in the forth semester so that I can go ahead with it and if it interests me even then, then I can pursue my career in data science and that's why I would recommend everyone here that mathematics is like the key for machine learning and deep learning, any kind of artificial intelligence techniques. During my master's, I also did part-time as a data scientist at Allianz. I was also getting some industry experience there and my thesis actually helped me realizing that I'm really loving this field and also as a data scientist, I knew I would be able to also focus on research and development so I wanted to pursue data scientist's job in the end after my masters. Then I got the opportunity to work at SAP, where I was now mostly focusing on researching different machine learning use cases for SAP and also developing them and producing them in the end. That's my journey.
Jake Samuelson: That's great. Thanks for taking us through that. Maybe I'll kick it over to Kaushik to tell us your journey into building data and ML systems.
Kaushik Rangadurai: I started with a master's in computer science at [inaudible] . I took a lot of ML courses there. Let's say around 2010 -2011. It was slightly more harder to directly get into an ML job from grad school. But then I started working as a data engineer or a data production engineer at Microsoft, where I was part of the determinant team. We were building data pipelines to understand the search query logs and stuff. Then moved into an ML infrastructure engineer at LinkedIn and then after about a year-and-half or so, switched teams and then join the recommendations team at LinkedIn and then since then I've been working mostly on Machine Learning then I worked on the search quality as a subtracting team at Google and then as an NLP engineer at Passage AI and then now look back to recommendations here at Facebook or Meta.
Jake Samuelson: That's great. Let's call it here different paths from the two of you. Frank, you're up next.
Frank Chen: Yeah, cool. Hi I'm Frank, as Jake said, my first job had nothing to do with machine-learning. I was a sophomore in college and then basically what happened was that I was working in Andrew Ang's lab once the summer, my freshman year. Then basically doing machine learning research but back then I was a freshman in college I didn't know anything about machine learning. What happened was that Andrew basically walked in and said, "I heard you guys know how to build a website so stop doing a research and help me build this website instead." That lead to me and a few of my friends in the lab joining Coursera when they started. I actually had not a lot of background in machine learning.
Frank Chen: At Stanford, both my bachelor's and my master's was in systems, computer systems, operating systems and compilers, and so on. After I left Coursera, I was in the machine learning field, this is back in 2016-2017 it was really beginning to blow up. So I decided, okay, cool, I probably should look for something along this field instead and so I decided- I didn't know a lot about machine learning, but at that point I knew something about systems, so in the end I found a position in Google Brain working on TensorFlow and TPUs, which are Google's AI machine-learning accelerators at that time. Because they are transitioning of Nvidia to their own chips. I worked on those things before here. That's how I got to be in machine learning. In particular, on the infrastructure side. I started spend a bit of time doing research, but a lot of my time was spent like making, say ResNet go faster or making Google's core machine learning or deep learning models go faster. So a lot of my time is spent on what's really now called machine learning engineering rather than machine-learning research. That's how I got to be in this field. Cool.
Jake Samuelson: That's great. Maybe it's really cool to hear your different backgrounds and passed into ML. Franken, systems engineering and Kaushik and applied Machine Learning and Shalvi, I guess in data science and different paths. But really, yeah, it's cool to hear that the different paths of letting you also be successful in different education trainings. Maybe I'll pivot on your grad school. For someone who's interested in ML, maybe I'd love to hear different perspectives. Do you need a PhD? How do you think about that? And maybe talk about your own grad school choices? Anyone? This is a popcorn question. Kaushik, you look ready to jump in.
Kaushik Rangadurai: I can start first. I don't think you need a PhD to work on applied ML industry. I would say quite a few folks in the industry, they don't have a PhD. We're working on applied ML. I think if you look at applied ML from industry you can break it down into five broad problems or tasks. I would say, even starting to work on any problem, you do some analysis or like an opportunities. I think that's the first part. Then you work on data preparation, trying to prepare your data set for training. That's the second part. The third part would be the modeling part where you go create a model to train the dataset. Then fourth would be evaluate against the baselines of production, and then the fifth would be experimentation. The reason I say that is, as you can see, the modeling or the ML part is probably one fifth of it and you would probably spend an appropriate amount of time in modeling as well. But there are about four or five things that cannot go hand in hand for you to build an ML like a product that's sneaks on the users. This doesn't even come into the online instructor and then linked applying and serving and stuff. I would say for anyone to have a firm understanding of various stages of a pipeline and then trying to iterate on the quality of each and every one of them. It's a necessity skill to excel as an ML engineer in industry.
Jake Samuelson: You just saved me some money. Thank you Kaushik. I don't know. Shalvi, I guess one thing you talked about the mathematics foundations and what are different pathways people can potentially get some of the math background to, obviously just a part of the job. But as you mentioned foundational for a lot of the work, like what are different pathways you've seen people to get that background?
Shalvi Mahajan: Yes. I understand that anyone can work on building machine learning models because we now we have Python library for that. So it's not a big problem doing that. But in order to understand what's going on inside our classifier let's say, we should know the mathematics behind at least I agree with that. That's why I would say so there are lots of YouTube channel from where I also made my foundations much clearer because in my bachelor's I was not in touch with mathematics so much except probability and so on. But then later I needed to get in touch with statistics and everything. For that, I usually used to watch one of the channel called Three blue, One brown from YouTube. That was really helpful for me at first and later on. If someone is not doing master's, they can also go with some
machine learning courses and then I think it's easy to follow.
Jake Samuelson: Great. We're fans of that YouTube channel as well. Hopefully I answered the question in the Q&A with that. Awesome. Let me get practical for a second. Maybe I'll kick it to Kaushik. Let's start with resumes. This is getting very practical. But some studies are saying like he was a hiring manager, maybe you're at Meta and you get tons of resumes and study say maybe you spent 15 seconds on the resume. Maybe that's true, maybe that's not true, but it's certainly, you're using some heuristics to look at things. What do you look for in figuring out if there's a match for your team in your Oregon, so people thinking about resumes and how to position themselves.
Kaushik Rangadurai: Yeah. In terms of skills, I would say a practical knowledge of applying ML to build products. This is something I would say most people in industry, after talking to a candidate for about half an hour or like 45 minutes, we'll be able to figure out if they've had some experience in solving problems. This is something that I feel you can't get from books, you can't get from a theoretical knowledge. I think while it's super critical to have that, I think once you get into the industry, one thing you will learn also to understand that there are quite a few things that are constantly changing. For example, when in about 2015, we were applying logistic regression or like boosted decision trees at LinkedIn to solve problems. But now, quite a few techniques have moved towards the wide and deep networks. Even to solve same problems that we were doing. During interviews, we don't look out for usually any particular techniques that they've used, but I think what's more important is the approach that they took. Why did they take the approach and what do they think they could have gotten better? I think to test their practical knowledge on solving problems and also what do they do when things don't work the first time, because this is often the case in the industry as well.
Jake Samuelson: That's great. Just to play back some of that, if you were preparing for an interview, maybe with you Kaushik instead of really thinking about some key examples when you had to build products and really just almost thinking through the case, like how did we go through. Maybe it's the five steps you said around like, is this a big opportunity? How do we model it? How do we evaluate? What roadblocks we've hit and how did we solve them? Is that some advice on preparing for an ML interview?
Kaushik Rangadurai: Yeah. I think having a good theoretical foundation is definitely necessary and it helps. But also, I would say, let's start solving problems. I mean, the projects that very consult them end-to-end and then try to iterate on those problems and then try to improve your accuracy or improve your performance in whatever way possible. I think even you will find that, let's say if you worked on the problem for about three or four iterations, there are sudden tips or techniques that
you learn that doesn't work for certain cases and things that you learn that works better for certain cases. Then that will also help you get a better understanding of the concepts as well and it will also help you in doing well in your job once you get the interview.
Jake Samuelson: That's great. How about you Frank when you think about as your interviewer you had on, and maybe you can speak to a little bit of what it's like to interview at a startup and maybe think about the matching of what people can consider whether they're better suited to thinking about Amazon and Facebook or Google or Nina. You've worked at both types of companies, smaller startups as well. Maybe think about the matchmaking process that's involved of trying to figure out where's the best place to take your career, smaller companies or bigger companies?
Frank Chen: Yeah. I really do think that people should have experience in both startups and larger companies. Those fundamentally they're very different businesses. If you are going to Google or you are going to Meta, or you're going to Netflix or Amazon, you're probably working alongside dozens, or your team that might be like dozens of people, so it's really really big especially if you are like entry level, if you are just graduated from your master's degree, from your college, you're probably going to be working on like a tiny bit of a much larger system. Whether it'd be like YouTube recommendations or whether it'd be even the stuff that I'm most familiar with on machine learning system learn. On TensorFlow chart, on PyTorch. You're going to be working on a tiny bit of it.
Frank Chen: But the iteration cycles will be a lot longer, but once you get a feature deliberate like a lot of people who use it or if you get even it's over one percent improvement in some metric on say, YouTube recommendations, that is, that actually improves the quality of recommendations for a lot of people. Iteration cycles are long. It's like you might encounter a bit more bureaucracy, a lot of meetings, a lot of approvals to actually get your things to launch, but on the other hand there's going to be a lot of impact from work just by the sheer scale of the products that these companies put out.
Frank Chen: On the other hand, in a startup, especially if you're looking to join an early-stage startup, there's still a lot of them in the machine-learning AI field. Iteration cycles will be a lot faster. Everyone is trying to get product market fit. Everyone is trying to at least get to cash-flow positive before they have to raise money or everyone cost of fundraising or so. It's a lot more hectic, but in some sense is a lot more fun, because you get to build a lot more things a lot quicker than if you were to actually then at a big company. I assume from the introductions, some of you are still in school, so you should probably now be actually a good time to actually just try out, just take an internship at a big company and then take an internship at a startup. Then hopefully that will help you figure out better being at a big company is more suitable to your personality and your lifestyle, being at a smaller company, being a startup it's more suitable.
Frank Chen: I think this is probably a good opportunity for someone who is already in the workplace. Obviously that's going to be harder, but even just coming to these events, just talking to your friends or working in larger companies, working in smaller companies will help you get a sense of what are working style is like, how busy they are and if they are working overtime and things like that. Yeah, just a lot of business if I talk to people and try to figure out what you want. I guess the other thing I want to say is that in machine learning, Kaushik and Shalvi has talked about the modeling aspect, the actual machine learning research part. But if you want to be in machine learning, for people like me with no academic machine learning or academic data science background, the way to do it, it's just, there is a lot of engineering requirements around getting machine learning systems into production. So if you just have a software engineering background, the best way to get into this field is just to start looking into job roles where it's just machine learning engineering. Because a lot of that is just being a good software engineer rather than I need to know a lot of math or I need to have published papers and things like that. I don't think I know a lot of math, I've never published a machine learning paper. But yeah, I was a Google brain. It's got a lot of opportunities all over for if you are actually interested in the field. Cool?
Jake Samuelson: Cool. I think that advice of working in adjacent roles or where you get a lot of exposure, by maybe you do something that you have more background in and you can add value right away but also have the balance of learning new things. I think pretty good advice for a lot of people moving into roles. Maybe Shalvi I'll go to you around the practicality of preparing for interviews and I know you're probably leaving grad school and joining the workforce and doing different things. You're spending time finishing your work and getting internships but also preparing. There's so many different skills required and there's statistics and engineering and math and machine learning background. How did you even think about spending and balancing your time for interview prep and all those different areas if you're meant to be fluent in several of them?
Shalvi Mahajan: Yes. It was really hard when I was doing my masters, I was also working like I said. I was working as a data scientist, Italian part-time. It was really hard to balance everything. Also making the foundation for the context and also working practically just like Kaushik and Frank said. Working on different projects in the university, working also on different challenges, for example, on Canvas, so I've to make my concept much clearer. But what I used to do was I used to do one thing at a time and do not think about other things as well. That really helped me actually focusing on just what you are doing and properly managing what you want to do. If you don't manage it, it will be totally chaotic. That's why it was really helpful for me to focus on one thing that I'm doing and then think about the other things when you
are doing it. That's how I could manage multiple things going on at the time I was doing my masters.
Jake Samuelson: Cool. Anyone else? Kaushik or Frank like thinking about how you prepared for new roles and interviews.
Kaushik Rangadurai: In terms of preparing for interviews, I would say, for most ML engineer roles you can divide your interview into a software engineering interview and ML engineering interview. I'm just going to focus on the ML engineering interview part of it. I'd say just slip brushing up on like some of the fundamentals because you might be in a bit rusty or you've not used them for a while just [inaudible] . Also looking at for example if you're entering at a company, let's say like LinkedIn or Meta, just going over the problems and then thinking through where MLS, like most used in this program. I might say if you take YouTube, I think YouTube recommendations it's like [inaudible] use case.
Kaushik Rangadurai: Then just think about, let's say if you are an engineer who is asked to solve the YouTube recommendations, what would you do? What are the problems that you're seeing in YouTube recommendations right now? How do you think you can fix it from the ML side? Then you can go with such as well like the YouTube [inaudible] . Just like brainstorming a couple of problems that are relevant to the company that you're interviewing for. Then just doing a background on the common techniques to solve those problems, and then also try to value your product has and see what are the common problems that as an end-user that you're facing right now in these products and then coming back to ML hat on how you can solve these problems from mission learning point of view.
Jake Samuelson: Great.
Jake Samuelson: A couple of questions coming in, so let me pay attention to those. One question from Rainier and thinking about the concept of reproducing papers. Maybe this idea of starter projects. Maybe you don't have an ML job right now, so you don't have the dataset and you don't have work that you can work on. What advice you'd give to people to start to get hands-on if they're not in their current role? I don't know Frank or Shalvi, anyone who wants to answer that.
Kaushik Rangadurai: Let's take a course at co:rise.
Jake Samuelson: I didn't plan it.
Kaushik Rangadurai: Also doesn't select Kaggle where there's a real datasets available and then you can go and practice problems and then submit to the leaderboards and click to evaluate yourselves. I think once you're more familiar, you can also go to existing papers and then see the datasets that they're using in those papers and then you can take the same datasets. The advantage of doing that is you'll have a good baseline on what are the metrics that they are evaluating on and then like what is a good baseline value for these metrics? It's just easy. These two methods, can I get into this baseline?
Jake Samuelson: Shalvi you unmuted. Were you going to say something?
Shalvi Mahajan: Yes, I could only agree with Kaushik. Kaggle actually is the very great platform where you have real dataset and challenge where you can solve problems and also reading a lot of papers and seeing their GitHub repository, there you will find a lot of how they approached the problem. Performance metrics does not need to be always accurate. It depends on the problem. Based on that, you can see how the paper that develop the performance metrics and how you can also work on it. It could be some business metric as well, depending on the company you are working in. But just introduce that accuracy is not always the perfect metrics. There are associated other performance metrics as well. The more you work on problems, the more you'll understand better I think.
Jake Samuelson: That's great. I'm going to take another audience question, an anonymous question. Is ML a young person's game? When you look around, you only see the young people or maybe talk about your work and think about if you have known any examples of people that have transitioned to ML later in their career that you've worked with? I'm curious to get that answer to that question. I'll toss it up.
Frank Chen: Yeah, I don't think so. I think in general, it's not a young person's game because for that matter I also don't think software engineering is a young person's game. Coming from a perspective of being a software engineer working in machine learning, this is not like competitive chess where you pick your bio mid 30s or something like that. It's not going to be anything similar. I think one of the things is that you do have to keep up, especially if you're in a research or if you are in a cutting edge stuff, you do have to keep up with new papers and things like that and that can take some more of your time.
Frank Chen: But that can take more of your time in machine learning, specifically, as opposed to say, systems, because in systems. The compilers haven't changed that much in the last 10 or 20 years compared to machine learning where a few years ago we were all talking about transformers, and now we're all talking about GPT-3 and so on and so forth. Lambda, all these are much bigger models. I think it's to just
keep up with the new developments in the field. It's going to be taking up more of your time and you just have to budget for it. I think almost all the researchers I know just spent a lot of time reading papers and re-implementing some of those papers and iterating on those papers. I think this just becomes part of their job description rather than anything else.
Jake Samuelson: Great. Thanks Frank. One question. I don't know. Kaushik could have answers to. What do ML engineers actually do? Maybe, is there the example that can bring to life what you spend all your time doing and getting paid what else to do? What does that mean? Give an example just so we can demystify it.
Kaushik Rangadurai: I can take that one. An example would be like YouTube. I would say that if you go to youtube.com, the order in which the videos are ranked, that's decided by an ML engineer or by the work of an ML engineer. The reason, it's so critical, let's say, if you didn't have them, then you can go with a bunch of Flickr heuristics say, I'm going to start a YouTube feed based on popularity or based on your past engagement. I think we have run numerous experiments to show that trying to optimize for something that the user likes. It can have through to produce huge gains to the metrics to the company.
PARAGRAPH: In terms of the actual test, at least for me here at Facebook or Meta, I would say, I can break it down into phi tasks. I think the first one would be in some opportunity is rising for me to start working on a new project. This is what my time
going to spend two months, three months. Then if it's been two or three months, what is opportunity in this project. That's the first part of it. The second would be the data preparation then the modeling part, then evaluation and then the experimentation and then followed by online infrastructure on the planet. These are the various subsystems within ML engineering if we can call it that I touch upon.
Kaushik Rangadurai: That's great.
Jake Samuelson: Maybe Shalvi while you're there you can talk about an example from your work just to eliminate a little bit like what is an exciting day in the life that maybe not an average day, but it's an exciting day in the life in terms of your role.
Shalvi Mahajan: In my work, basically we are mostly focusing on validating machine learning use case. During the day there are lots of brainstorming sessions that we do. Then when we come up with certain approach, how to approach this machine learning use case. First, we validate whether we even need a machine-learning in it or not? If you think there is a requirement than we think of how should we approach it and how should we build the model for it. Once the model is finalized, then we also do some business value assessment, some different performance metrics
comparison. Then with regards to the some baseline that we have. Then finally, we determine whether this machine learning model should go in production or not. That's all what I do in my job and a day looks like lots and lots of brainstorming sessions and meetings. Some obviously development and research included.
Jake Samuelson: Great. Frank or someone to talk about just to ML upside. Like it's obviously becoming a bigger and bigger part of the ML ecosystem and there's more jobs and maybe these titles mean different things at different companies. But there are soon to be a growing set of MLOps titles. But maybe thinking about a starting point of like, what does the MLOps do? Then maybe a starting point of how someone could learn more about the ML production as well.
Frank Chen: Because there are multiple roles in this industry. You have like machine learning research or machine learning scientists like Google is called. These roles are usually called research scientists roles. Those are the typical roles that you think about Google Brain or you think about FAIR, for instance, when you think about parts of open AI, those people publish papers and basically they come up with GPT-3, or DALL-E, or PALM or any of the large models. This is what they do is they are similar to enroll. My grad students or professors at academic research labs, but instead of having grad students do all the work, you have lots of CPUs or GPUs to do the harder work instead. Those are research scientists. Then the second role is machine learning engineering, and this is a very big spectrum of roles. Some machine learning engineers will work closely with the research scientists to implement the new models that they have come up with. Because oftentimes these new models for instance, like PALM is something like 540 billion parameters. Sometimes they will require a lot of engineering to get to work. Because you're not going to be able to train any of these models on one GPU or one TPU. A lot of distributed systems problems come up when you, and sometimes these engineers get their names onto the papers as well because they have contributed so much to just getting the models to work to demonstrate that they actually are better. Then some other ML engineers would, are more responsible for things like productionizing existing models. Because for instance, YouTube recommendations, or search ranking, or ads ranking, a lot of these models are not public, but you still have people who need to make sure that the data ingestion pipeline works, to makes sure that the training works, to make sure that you can still serve these models, because all these models if you just don't optimize the survey, is just going to be really expensive or really slow. A lot of machine-learning engineers also work around this. Then these roles all tie into the largest DevOps, and SREs, and just normal software engineering practices. Those are separate roles. Then when people talk about MLOps, they are more referring to the second set of people rather than the machine-learning engineers who are working in research scientists. Basically Ops just means ope rationalizing the ML models so that you can actually get business benefit out of basically using ML and not have models that don't work, or models that cause crashes, or models that take three seconds to come up with results that your deadline is 0.5 and so on and so forth. A lot of like just making sure that models
work and the models work well, falls under MLOps and machine learning engineering.
Jake Samuelson: That's great. I know it can seem a little bit like alphabet soup and talking about different roles and different titles and different companies. But I think that was quite helpful and laying out the spectrum and different types of roles, things. Thank you, Frank. Maybe I'll take a question from Rangadurai, I guess the question is, which areas of ML and deep learning think is maybe easier to get started with. It sounds like some of the advice is to follow your interests and follow what's interesting to you. But if there's a practical way to think about, is there computer vision easier to get started for something completely new or something else? Maybe if you could address someone could address that question.
Kaushik Rangadurai: Well, I think the good thing right now is there is a lot of convergence across fields. What I mean by that is like the techniques that you learned in one field is probably not applicable only to that field. Transformers started within the NLP field, but then it's widely used in computer vision as well through vision transformers, for example. The same way like the conditioning, which started as a Computer vision. That concept is also widely used in NLP, these days. From a skills or knowledge perspective, I think you're not limited to just the techniques that you've learned in one field. I also find, like search to be an easier field to get started with one. That is, I would say like searching Computer vision would be easy fields to get started with. Like search because you can, there's a short query that's probably easily understandable and then you can go and look at the results and it's easy to debug and get right on that.
Jake Samuelson: Great. Thank you so much. Well, I'll do one more plug to co researchers, a great search track. What we found is that a lot of people working in Search they didn't have a master's degree in search. There hasn't been a whole lot of trading and to connect with experts in Search and to connect with other people. Maybe working in those roles at other companies. It's a really cool experience and I think really very high return on your learning investment. Because you can, not only are their roles available, but you can come back to your company and really add a ton of value if you can improve some small things with their search system. To second that. Great. Well, we're actually going to split out Judy, is that right? I think we're going to split it in a few minutes, but there was a couple of questions that we didn't answer, but I'll figure out a good way for us to all connect. Frank, there was a specific question around preparing for Google Brain and interviews. I'll let you guys have some fun on that. But maybe one last question I'll have you guys answer is like, it's so much is happening. Frank mentioned some of this, but how do you keep up with trends and what's your news feed look like or how do you spend time in the day? Pick your head up from the urgent needs of the business. This pay attention to what's happening and what you can maybe bring it back to your work. Shalvi, I don't know if you have thoughts on that question.
Shalvi Mahajan: Basically, I really like subscribing to certain news, let's say for big companies, Google or Meta. Thereby I read what's going on in the technology. I am also connected with SAP blogs or with Corise anyways, we are providing so many good courses that are very much related to what the recent technology is about. In my job description the work I'm doing there also there, we need to study in lots of paper and we are always focused on mostly the new technology and not the old ones, which due to some reason didn't work out. That's how at least I keep myself updated.
Jake Samuelson: Frank, you always have a couple of monitors out there one time.
Frank Chen: I think a lot of it might be like following Twitter and things like that, just following researchers that you might, people post interesting things or retweet interesting things. Obviously, I couldn't possibly going to archive and read all the papers. But sometimes just skimming what other people are saying we will actually be helpful and it actually comes out of my work in the future who knows. I can always go back and read it.
Jake Samuelson: That's great. Well, Judy, you're there? I think.
Judy Zhu: Yes. I couldn't find the mute button but I'm here.
Jake Samuelson: Great. Yeah. I think where I could pack another hour into this, but I think we're at time. Like I said, I'll make some of these connections, but thank you so much for the questions and you made my job easier as a moderator to have really useful questions in Kaushik, Shalvi, and Frank. Really appreciate your time and to really help people think about the different options and how to transition from where they are to an ML career. Thanks for taking you're time. We really appreciate it.
Judy Zhu: Thanks so much you three. This event is the first part is public for everyone, but now we're going into the private part, which is, we do have our current ML track students and we're going to give 15 minutes to them for private breakout with the three panelists. If you are a current track student and I see some folks here, Emily, Eric, Scott, Josh, Maggie, you can stay too, you took two of the three courses. Maggie took one of the courses in track, loved it so much and then bought the rest of the track. You know who you are if you're in track, please stay and then the rest of you, thank you so much for coming and we'll share out a recording of this first part soon. Thanks everyone.