70% of machine learning engineering candidates I interviewed failed in the same round, and the thing is it's actually really easy to pass. All you need to know is what you're doing. Having worked at companies like Adobe, Twitter, and Meta, I spent the last 10 years on

both sides of this interview and realized it all comes down to acing these six things. So in this video, I'll reveal what kinds of questions I had to answer to land a job at Meta, share my system for answering any machine learning design questions your

interviewer might throw at you, and I'll even give you one stellar strategy to prepare in less than a day. But before I can share my system with you, you first need to realize that you're preparing for your interview all wrong. If you're spending the hours leading up to your

interview reading up on loss functions, then you're wasting your time. The thing that trips up most candidates has nothing to do with their ML knowledge. I played a lot of Microsoft Flight Simulator when I was in college. I know where every button and instrument gauge

on the Cessna 150. Recently, I had a chance to take the yoke of an actual plane and, uh, well, my friend had to take over to keep us both safe. The things that keep you from landing your dream job are not about the buttons but the feel of the

interview. Specifically, insufficient understanding of the question, confusing your interviewer, and poor time management. And that's where my system comes in. Using the system allowed me to prepare for my interview at Meta in half a day, and the interviewer was so impressed that I got an up-level from

The senior position I was interviewing for, to staff interviewing at Fang companies, especially at senior levels, is like shooting for the moon. Say you were actually tasked with that, I give you unlimited funding and say, "Fly to the moon. It's up there." And vaguely point up

in the sky. This is sort of what the ML design interviewers do. They give you a generic question and then sit back watching your spacefaring dreams sail by. Being off by even half a degree on your course to the Moon already assures you will never get there if you don't

fully understand the problem. You might overcomplicate the solution or fall short of addressing it fully. Make sure you take 5 minutes or so to really understand the question. You sound a lot more competent when you make assumptions than when you pester your interviewer

with questions. So limit the amount of clarifying questions you ask. And instead of panicking about details like exactly how many users does Facebook have, just say "Facebook has about 3 billion users." This next stage should take you about 6 minutes, but if you neglect it, you will

spend much more time in back-and-forth explanations with your interviewer. Sadly, too many ML candidates fail because they use this stage as the opportunity to dive deep into rabbit holes, showing complete inability to do the high-level design and wasting precious time, making it

impossible to recover. Instead, you need a rocket blueprint. You need to be very deliberate about building a very high-level system diagram with barely enough detail to address the main requirements of the problem. So instead of XGBoost, you need to think "model" and leave the

specifics to a later stage. Doing this R avoids most miscommunications between you and your interviewer. Be familiar with whiteboarding software you will use in the interview. You do not want three out of six minutes in this stage to go to figuring out how to draw a square.

While having a good plan goes a long way toward clearing this interview, this next part is the second largest component of your score. Everything you have done in the first 11 minutes is a great setup for a successful interview, but this stage is where ignition happens, the

rocket lifts off the launchpad. The objective of the next stage is to pick up speed and not experience a rapid unscheduled disassembly. I know I have experienced a few of those in my interviews over the years. How about you? Let me know in the comments if you have

done the pre-work. This is the fun part, of course. I'm talking about data considerations, and there are a lot of data questions to consider. To get this space, you must at least understand what your labels are and where they will be coming from. Understand what your

features are and how to translate them into numbers. Understand how to normalize those numbers. Understand how to split your data set. Are your data imbalanced? They almost always are. How will you address that? This part should take you about 8 to 9 minutes. Note that this is a lot of

ground to cover in a short amount of time. You should not discuss every feature, just a few representative ones. Congratulations, you're in low Earth orbit, but this next boost will get you all the way to the moon. This section covers the most important thing your interviewer is

Looking for the next 15 to 16 minutes are critical. This is no longer time for clarification; you need to make the most out of your time. Your task in this stage is to talk about modeling metrics and training, as well as overfitting and

other issues that may come up, like cold start and time travel problems, and propose solutions to them. This is where you demonstrate your machine learning knowledge targeted to the scenario your interviewer actually cares about. While this is a lot of material and you may be asked a wide range of

questions, don't panic. At its core, almost all of machine learning is either about classification or regression, and in fact, many problems can be solved with either. Once you define your problem as one of those, there are standard paths that will keep you safe. Congratulations, you're now as

successful as Apollo 13. Wait, no, that's a bad one. Apollo 8, both were piloted by Jim Lovell. Anyway, Apollo 8 is a pretty good comparison here, though. It was a test mission that went around the Moon without landing. Earlier in the Apollo program, if you're early in

your career, you may already have it in the bag, but just to be sure, or if you're looking at more senior jobs, let's land on the moon. Stage five: you have 5 minutes to knock the interviewer's socks off. Your one small step... interviewers will

sometimes focus on how to launch and operate a machine learning system in production, but if you are an expert in training parallelization or online evaluation techniques, this is your time to shine. There are often multiple ways to solve the problems you will find in this section. At least once, you should come up

with trade-offs between two or more reasonable solutions and make strong recommendations for one of the wait I thought you told me there were six stages indeed while the rest of the world was rooting for Neil Armstrong to step on the moon I am sure his family

would not have liked the mission as much if he didn't land safely stage six is just for you it would be a big bummer to break into the company that breaks your will to be in the industry so use the remaining time to ask the interviewer

questions about working at the company you only have time for one or two questions here so make them count if you follow these steps with more or less suggested timing you will ace your ML design interview no matter what question is asked but you want to know what

questions are usually asked don't care well I have some good news for you the great majority of the questions for machine learning design interviews come from just a couple of areas of machine learning and even if these are not your area you can get up to speed on these by

reading just a few papers and blog posts Linked In the description most big tech companies focus on only two types of questions for their generalist positions these two question types are popular because they test a broad swath of machine learning knowledge and the interviewer can dive into a lot of

places to understand what you know the second most common type of question for this interview is harmful content or inappropriate content something like detecting firearm in an Amazon listing or nudity in a social media post this is the type of question I got asked that

sealed the E6 meta offer for me, but even more common is the question I asked as an interviewer in big tech: design a recommender system. Something like a Twitter timeline or Amazon product recommendation. For specialized roles and in smaller companies, the questions tend to be very targeted to what your role will be. If you are a vision specialist, expect to design a vision system. While machine learning is a very deep field, a handful of areas are over-represented in ML design questions. However, you should never assume that you will get one of those problems. Instead, let's talk about

how you would study so it doesn't matter what questions you get asked. The machine learning system design round is one of the many interviews you're preparing for on top of all the other things you're doing. How do you prepare for it fast and without sacrificing sleep? Luckily, in

addition to my experience applying to Meta, I have since helped dozens of people get into FAANG companies at levels from new grad to principal engineer with this advice. And you already know the first step from earlier in this video. There's an apocryphal story of Charles

Proteus Steinmetz, a renowned electrical engineer who was called to fix a generator for Henry Ford. After marking a spot with chalk and instructing the engineers to make specific adjustments, he charged $10,000 - $1 as he said for the chalk mark, and $9,999 for knowing where to place it. Understanding the six stages of the interview is the absolute prerequisite to doing this efficiently. Pick one of the questions I mentioned in this video and go through how you would answer each stage in brief. If you know offline metrics really well,

just check them off your list this is your $9,999 task figure out where to place your efforts then you read up on topics you identified as weak spots utilize resources in the description when you need them skip the ones you already know at this point you have learned to swim

by reading a book well remember when you tried to learn how to swim using the internet I did learn how to swim on the floor the skills are transferable I just have no interest in going in the water Kudos however a real pool full of water may present some

different challenges whatever else you do plan to do two full mock interviews you can do a paid mock just be very careful to look through reviews or just buddy up with a friend who is at a similar level I will walk you through how to conduct a mock interview in a

future video on this channel if you apply what you learned in this video do a few hours of additional research to fill the gaps in your understanding and do two mock interviews you will be perfectly positioned to ace your ML system design interview getting through the ML Design Round is only part of what

it takes to have a stellar career in Big Tech to learn more about what it takes to be an MLE hit subscribe so you won't miss my future videos

This ML Design Interview strategy got me into Meta

MLEpath

머신러닝 엔지니어링 면접에서 흔히 저지르는 실수

응시자의 70%가 쉽게 피할 수 있는 실수로 인해 불합격합니다.

성공은 6가지 핵심 영역을 마스터하는 데 달려 있습니다.

사소한 부분(예: 손실 함수)에 집중하는 것은 효과적이지 않습니다.

핵심은 기술적인 세부 사항뿐만 아니라 인터뷰의 '느낌'을 이해하는 것입니다.

중요한 문제는 질문을 이해하고, 혼란을 피하며, 시간을 효과적으로 관리하는 것입니다.

면접을 잘 치르기 위한 6단계 시스템

반나절 만에 메타 면접을 준비할 수 있는 시스템입니다.

ML 디자인 인터뷰는 모호하게 정의된 우주 미션에 비유할 수 있습니다.

질문을 이해하는 데 시간을 할애하고, 명확한 질문을 제한합니다.

처음에는 불필요한 세부 사항을 피하면서 개략적인 시스템 다이어그램을 작성하세요.

세부적인 내용을 살펴보기 전에 높은 수준의 시스템 설계에 집중하세요.

데이터 고려 사항(레이블, 기능, 정규화, 데이터 분할, 불균형)은 매우 중요합니다.

모델링, 메트릭 및 잠재적 문제 해결

과적합 및 기타 문제를 해결하면서 모델링 메트릭과 교육에 대해 논의하세요.

이 단계에서는 시나리오와 관련된 구체적인 ML 지식을 보여줍니다.

분류 또는 회귀를 사용하여 많은 문제를 해결할 수 있습니다.

프로덕션 시스템 고려 사항 및 장단점

프로덕션 환경에서 ML 시스템을 시작하고 운영하는 것에 대해 논의하세요.

교육 병렬화 또는 온라인 평가에 대한 전문성을 강조하세요.

서로 다른 솔루션 간의 절충점을 제시하세요.

회사에서 일하는 것에 대한 통찰력 있는 질문을 해보세요.

대부분의 질문은 머신 러닝의 몇 가지 핵심 영역에 속합니다.

빅테크 기업은 일반적인 질문(예: 추천 시스템)에 집중합니다.

유해 콘텐츠 감지는 또 다른 일반적인 질문 유형입니다.

소규모 기업의 전문 직무에는 더 구체적인 질문이 있습니다.

6가지 면접 단계에 대한 이해를 우선시하세요.

취약한 영역을 파악하고 그에 따라 학습 노력을 집중하세요.

두 번의 전체 모의 면접을 실시합니다.

This ML Design Interview strategy got me into Meta

This ML Design Interview strategy got me into Meta

스크립트