머신 러닝 학습: 엔지니어 지망생을 위한 필수 조언

### ML 엔지니어 되기: 학습과 성장의 여정
* 🚀 6년간의 헌신과 학습 끝에 꿈에 그리던 **ML 연구원이라는 직업**을 시작하게 되어 기쁘고, 머신러닝을 효과적으로 학습하는 방법에 대한 저의 인사이트를 공유하고자 합니다.
* 📚 선형 회귀, 서포트 벡터 머신, 랜덤 포레스트와 같은 고전적인 ML 모델의 기본을 파악한 후 **고급 주제나 생성 AI 도구**에 대해 알아보는 것이 중요합니다.
* 🔍 정확도, 리콜, 혼동 행렬과 같은 주요 메트릭을 이해하는 것은 숙련된 ML 엔지니어와의 인터뷰 및 토론에서 **기본 지식을 입증**하는 데 필수적입니다.
* 💻 머신러닝은 탐구와 실무 경험이 성장에 필수적인 경험적 과학이므로 **실제 프로젝트를 일찍 시작하는 것**이 좋습니다.
* 🛠 강력한 소프트웨어 엔지니어링 기술은 ML 직무에서도 마찬가지로 중요하며, **이론에 집착하기보다는 프로젝트 구축에 집중**하면 실무 능력을 향상시킬 수 있습니다.
* 📈 매달 수천 개의 ML 논문이 발표되는 상황에서 잡음을 걸러내고 **팬더와 같은 기본 라이브러리에 집중**하는 동시에 필요에 따라 새로운 도구에 적응할 수 있어야 합니다.
* ⏳ 머신러닝을 숙달하려면 시간이 필요하며, **'10,000시간 법칙'에 대한 믿음**과 함께 시간이 지남에 따라 꾸준한 연습과 프로젝트 작업이 전문성으로 이어진다는 점을 강조합니다.
* 🎉 궁극적으로 학습 과정을 즐기고 도전을 받아들이면 수년간의 노력과 인내 끝에 **제가 그랬던 것처럼 성공적인 ML 엔지니어**가 될 수 있는 길이 열릴 것입니다.

성공적인 ML 엔지니어가 되려면 고전적인 모델을 마스터하고 실무 프로젝트에 참여하는 것이 필수적입니다.

So I get a lot of questions on how to learn machine learning and become an ML engineer or data scientist. I will soon start my dream job as an ML researcher, having been working towards this goal for the past 6 years and have learned a lot

about what to do and what not to do, at least in my opinion. And I've also seen a lot of advice from other people online that I think don't put that much thought into what they say. So in this video, I thought I'd share my honest advice on

how to actually learn machine learning. This is one of the most important pieces of advice I want to give because more and more people just ignore it. Let's be real, many of you will probably have seen ChatGPT and the other generative AI tools like image generation models like Midjourney or

DALL-E. And I get it, it is really super cool, but directly wanting to jump to generative AI is only really a good idea if you simply want to build a product that calls, let's say, the OpenAI API. And there are amazing online courses that

focus only on generative AI, but if you really want to become an ML engineer or researcher or something like that, that won't be enough. You need to learn the basics, and by basics I here specifically mean the classical ML models like linear regression, support vector machines, and

random forest. And classical ML know-how, such as what precision, recall, F1, confusion matrices, R squared, and so on are. You see, in the most practical sense, you will simply need that knowledge for interviews, but what I think is a really important indicator for skill is when

doing ML small talk with real ML engineers you might casually talk about these techniques and if you don't know them people will know you are not really that educated yet at least I know the old school ML is not the fancy stuff you see today but it definitely is core

knowledge a real ML engineer needs to have and is still used more often than you think so make sure to always have as much fun as possible but definitely don't skip these fun fundamentals you've had enough machine learning courses I know this might sound weird now but please let me explain what

I really mean by that being able to understand machine learning has some prerequisites like math and programming The logical next step is to watch some ML courses for example those by Andrew Ng on Coursera in those courses you learn the basics of ML and deep learning

which I mentioned in my previous point and get some practical coding experience so what now is the question what is the next course you should watch there are so many well honestly at this point I don't think you should focus on specific courses you should start to work on real

projects as early as possible machine learning is a very new and in most parts very empirical science you have learned the basics and at that point there is no clear path from A to B it's all exploration you will most likely feel lost that nothing makes sense that you

are wasting time and not progressing in the most efficient way possible but believe me you are doing just fine this is normal ML is still a really young field with so many separate developments there's a huge space of ML knowledge and you will be learning

different topics at random, you might learn about language modeling, which is somewhere in the top left corner of the ML knowledge space, and then reinforcement learning, which is somewhere in the bottom right corner. And there is no connection yet, but over time, if you just keep going and going, you

will see all the connections, more and more. Of course, you can never be an expert in all of them, but by simply staying curious, working on projects that involve these different techniques or domains, and perhaps watching introductory lectures on those topics, you will get a good, high-level

understanding of what is possible with each technique and what isn't, and you'll get a feeling for what you enjoy doing. But most importantly, you will get more practical experience with every new project. You will realize how bad you were in the previous ones and get better

and better. This process is very important, so don't waste time looking for the perfect next best course for too long after the basics. I honestly recommend for a more project-based learning approach because you are a software engineer. Pretty much all beginner ML students have a wrong idea of what doing

ML or working in ML as an ML engineer really means. Even ML researchers need to have strong software engineering skills. This took me a while to understand as well. Greg Brockman, a co-founder of OpenAI, said it's easier to teach people ML skills than software engineering. Another

great illustration of what the impact of software engineering compared to actual ML code and knowledge is. This visualization by Google showing how much the explicit ML code is compared to everything else that is part of an ML project. Again, ML knowledge is definitely mandatory, but as I mentioned in my

Previous point, don't obsess too much over ML theory after the basics. Try to build stuff. Don't think of these projects as a project to look cool on your resume, but as a project to learn. I.E. to actually improve your skills. And after a few such iterations, your

projects will naturally get better and better, more complex and interesting, and more impressive on your resume. And make sure to somehow show your work, for example in the form of a blog post or a working project website with a demo. All the debugging you will need to do, all the different data processing and

evaluation pipelines you will build, all the data visualizations you will create, the efficiency improvements and so on will be a large part of your project. And those parts will be equally as important as the few lines of actual Python code you will write, if not more so. Don't think

all this time spent coding and debugging is wasted. It is very, very important. This one is a scary one. There are thousands of ML papers published per year, wait, wait, no, per month. And there are always new libraries that claim to be better than others. How can one keep up with all the speed of

progress and information? Well, you need to realize that there is a lot of noise. It really isn't about learning all the newest libraries. Sure, PyTorch is cool, but by the love of God, especially as a beginner, don't get overwhelmed by such new libraries. You need to know a bit of

pandas, which almost everyone is still using. And if you will need to work with PyTorch at your new job, well, you will very easily adapt. I'm not hating on PyTorch by the way. I'm just saying you don't need to instantly jump to the

newest stuff and ignore the more important things. I mean, for my new job I will most likely need to work with Jax instead of PyTorch. I've never used Jax before and my interviewers knew that, but they know that I will be able to adapt

to the library. The basic libraries are enough and it even really isn't that important whether you started with TensorFlow, PyTorch, or Jax. It also really isn't about having read every new paper. It's more about recognizing trends and techniques that are being used more

and more and might be useful for your work. For example, especially for an embedded engineer, efficiency techniques are a big thing. For instance, at a certain point it will likely be important to notice the release of something like flash attention and understand what it is, or something like

KV caching or grouped query attention or something that currently is all the rage, like retrieval augmented generation, or for short, RAG. And if you haven't noticed, all these things don't really have much to do with new ML techniques, but those are engineering techniques. The amazing thing is that all this stuff is

so new that learning it will literally mean you are one of a fairly small group of people who are familiar with this technology. I know it might not feel like it because it might seem so popular, but think about another software industry like web development. Effectively, all the skills

you can learn there have been learned by many people who have started with web development perhaps 10 years ago and are now 10 years ahead of you. And since RAG and all these techniques are only 2 years old or so, you will be able to

catch up to the best much quicker than I said. Keep in mind the point I mentioned before with respect to projects: you need to actually get hands-on practice and work on projects to learn the skills and then be able to prove that you have the skills by showing your

projects, but all this it takes time. This is the most important point, period. So many other expert influencers try to sell a "become a Gen AI expert in only one month" product. What does "expert" even mean? The point is that simply won't work, at least not without a significant amount of

prior software engineering experience and a good understanding of basic ML. All of the points I mentioned before won't happen over the course of a few weeks, of course. How long it takes for you specifically to learn machine learning to a level where you are actually really

useful to an ML team simply depends on the level with which you begin this journey. Andrej Karpathy is a big believer in the 10,000-hour rule, and I am a big fan of it as well. If you invest 10,000 hours

into some skill, you will become very good at it. So don't expect to take one course, do one project where you train your first model, and then be ready for a real-world job. You need to work on project after project after project. Read into some literature whenever you

encounter something new, and repeat that pretty much forever. I know this might sound harsh, but I definitely don't want to discourage you. Rather the opposite: I want to encourage you to really learn machine learning, enjoy the learning process, and become a great machine learning engineer because I am certain

that if you enjoy it, invest the time and push through the bad times, you will become great at it and it is simply so fun and worth it on so many levels. I mean, for me it took six years of going to college, working as a part-time student

researcher in ML for over 2 and a half years, working on countless projects and failing a lot to now actually get my dream job as an ML researcher, which I still can't really believe. So if you want to really see what I went through

in all these six years, you might want to watch this video next. Bye-bye.

My Honest Advice to Beginner ML Students for 2025

👨🏻‍🎓 I like to train Neural Networks and understand intelligence
👨🏻‍💻 I talk about how to learn Machine Learning, how to get a job, and life as an ML researcher

Boris Meinardus

머신러닝 학습과 ML 엔지니어/데이터 과학자가 되는 방법에 대한 많은 질문이 있습니다.

자주 무시되는 기본 사항을 중심으로 머신러닝을 배우는 방법에 대한 조언을 제공합니다.

기초 지식 없이 제너레이티브 AI에 바로 뛰어드는 것을 경고합니다.

고전적인 ML 모델(선형 회귀, SVM, 랜덤 포레스트) 학습의 중요성.

정확도, 리콜, F1 점수, 혼동 행렬 등과 같은 개념에 대한 이해의 필요성

프로젝트 기반 학습 및 실무 경험

ML 과정을 시작하기 전에 수학 및 프로그래밍과 같은 전제 조건을 이해합니다.

기본 사항을 학습한 후 강좌에서 실제 프로젝트로 초점을 전환합니다.

ML의 탐구적 특성을 수용하고 길을 잃었다는 느낌을 받아들입니다.

다양한 ML 주제를 무작위로 학습하고 경험을 통해 시간이 지남에 따라 연결합니다.

실무 경험을 강조하는 프로젝트 기반 학습 접근 방식을 권장합니다.

머신 러닝에서 소프트웨어 엔지니어링 기술의 중요성

많은 초보 ML 학생들이 ML에서 소프트웨어 엔지니어링의 역할에 대해 오해하고 있습니다.

ML 연구자에게도 강력한 소프트웨어 엔지니어링 기술을 강조합니다.

ML 프로젝트에서 비ML 코드의 상당 부분을 보여주는 시각화입니다.

채용 담당자에게 깊은 인상을 남기기보다는 기술 향상을 위한 프로젝트 구축에 집중하세요.

ML의 빠른 발전을 따라잡아야 하는 과제를 해결합니다.

새로운 라이브러리에 압도당하기보다는 판다와 같은 기본 라이브러리에 집중하세요.

모든 새로운 도구를 마스터하는 것보다 새로운 라이브러리에 적응하는 것이 더 중요합니다.

모든 신문을 읽기보다는 트렌드와 기술을 파악합니다.

플래시 주의 및 RAG와 같은 엔지니어링 기법에 대한 이해의 중요성.

별다른 사전 경험 없이 빠르게 AI 전문가가 되겠다는 생각을 반박합니다.

유용한 레벨에 도달하는 데 필요한 시간은 시작 지점에 따라 다릅니다.

스킬 숙달을 위한 10,000시간 법칙을 강조합니다.

프로젝트, 독서, 반복을 통한 지속적인 학습.

학습 과정을 즐기고 도전을 통해 인내심을 갖도록 격려합니다.