HR/interviewers from multiple AI companies revealed: How we recruit machine learning engineers

Guo Yipu from Aofei Temple
Quantum Bit Report | Public Account QbitAI

Let me tell you a horror story. There is only half a month left before the end of the Spring Festival holiday.

In other words, half a month later, you can choose the company to switch to next year.

An interview sharing post has just become popular on Reddit. Many machine learning practitioners or technical leaders and HRs who recruit people have shared their personal experiences in interviewing machine learning technicians.

From here, you can see some long or short interview processes, some tricky or concise questions; you can also know why you failed some interviews from the perspective of HR and technical leaders who recruit people.

We have summarized the four major routines of machine learning interviews.

Routine 1: Use actual problems as test questions

Shared by netizen p-morais

Company size : Startup

Online/offline : mobile video interview

Stages : 2 or more

1. Candidates will give a half-hour to one-hour introduction of their research or work;

2. A 45-minute unstructured interview will be conducted by members of different teams. The main content is to give a problem that may be encountered in actual work and ask the candidate how to solve it.

Time : more than 2 hours

Evaluation : Suitable for startups/suitable for recruiting machine learning scientists/preferably supplement math basics test

Netizen p-morais:

"I like this kind of startups, because they can judge the suitability and ability of candidates for the position from their job descriptions. Because I am also a startup here. Maybe big companies like to hire smart newcomers and train them from a blank slate, but I can't hire someone just because he is smart."

Netizen drrelyea:

“It’s hard to find the right person this way, but once the right person shows up, you’ll know that this is the person you need to hire. However, for non-startups, you have too many positions to fill, and it’s too late.”

Netizen snendroid-ai:

“This is the best way. What better way is there than asking candidates how they would solve a problem you actually encountered? Your team has already spent a lot of manpower and resources on this problem and is sure to be familiar with all the solutions, so just asking questions in this area will easily judge the candidate’s ability.”

Netizen mydynastyreal:

"My company is like this too, but we also have a whiteboard session that tests all the basic math content from statistics to calculations. It's very stressful, so candidates will perform worse than usual. But some people who look very good perform very poorly. I would rather hire someone who is good at math and doesn't know much about AI than someone who has no math foundation."

Netizen AchillesDev:

“This is also the case when we recruit machine learning (CV) scientists. When we recruit engineers, it will be more like a typical coding interview, and will focus more on the overall design.”

Routine 2: Use homework to examine the candidate's coding level

Shared by netizen drrelyea

Company size : Undisclosed, but from the description it should be a large company

Online/offline : online first, then offline

Sessions : 2~4

1. The first stage is a telephone interview.

First, we spend half an hour to understand the candidate's research focus, experience, work initiative and other information. Then there is a 20-minute code interview, which will screen out about half of the candidates. After that, the candidates are given 10 minutes to ask questions to the interviewer.

1.1. Additional questions for PhD students with zero experience.

For PhD students who have no practical work experience in machine learning, we will send them an email after the phone interview, giving them some time to study some machine learning courses on Coursera, read Hastie and Bishop's books, and learn TensorFlow, PyTorch, pandas, and numpy, giving them two or three months to learn these things.

2. In the second stage, some positions will have homework.

The assignment takes about 3-4 hours of work, and examines whether the candidate's code is concise, can solve the problem, and can explain the code clearly. At this point, about two-thirds to three-quarters of the candidates failed because they did not understand machine learning or wrote poor code.

3.The third stage is the offline interview.

The 3-4 hour offline interview mainly examines the candidate's knowledge and understanding of machine learning, coding ability and overall quality, as well as open-ended questions based on the candidate's resume. The purpose is to ensure that the candidate's technical ability and knowledge reserves meet the requirements, that there are no problems with his or her character, that he or she can work under high or low pressure, and that he or she can handle both clear and ambiguous work.

The majority of candidates who fail at this stage do not have problems with hard skills, but rather overall quality issues, such as a lack of concentration on work, confusion about requirements, lack of initiative, being too arrogant, poor communication skills, etc.

Time required : 4~8 hours for experienced people, 2~3 months for inexperienced PhD students

Comment : Very helpful to candidates / Too much homework

Many people think that offering courses to candidates with basic knowledge but no experience so that they have time to study is a very conscientious act and will have a positive impact on a person's career.

Regarding the issue of assigning homework, netizen neal_lathia commented that the process of assigning homework will increase the burden on candidates, but having homework can provide more information than empty talk and can allow for more discussions with introverted candidates.

Routine 3: Interview and exam are all included

Shared by MeatIsMeaty

Company size : Undisclosed, but it should be a large company based on the process

Online/offline : first online, then offline, and finally take the test online

Sessions : 9

1. Phone interview

2. Behavioral Interview

3. Group Interview

4. Technical interviews (machine learning, statistics, coding, data engineering)

5. Speech interview

6. Modeling work

7. Analytical homework

8. Online code testing

9. Online Statistics/Machine Learning Test

Time taken : Undisclosed, but it looks like a long process

Rating : Too complicated

The contributor of this routine has now become a recruiter after going through such a long series of interviews. The phone interview lasted only 1 hour, and the offline interview only had a 2-hour technical interview and a 1-hour behavioral interview.

Routine 4: Ask questions on the spot

Netizen liqui_data_me shared his internship interview experience

Company size : Fortune 500 company

Online/Offline : Offline

Sessions : 2

1. Talk to HR.

2. Do questions on the spot, from 8 am to 1 pm, there are a total of 3 sessions, each session is LeetCode + machine learning questions + machine learning project experience + brainstorming to solve the problem. There is even a question about deriving the back propagation of a neural network.

Time required: more than 5 hours

Finally, from the interview and recruitment processes of different companies, we can see their emphasis and needs on talents. Some companies pay more attention to whether candidates can solve practical problems, some companies pay attention to technical capabilities, and some companies pay attention to people's comprehensive qualities.

Moreover, there are many companies with complicated processes, but there are also many interviewers who advocate simplicity and efficiency. This also reflects the corporate culture of the recruiting company.

I wish good luck to those who want to change jobs after the New Year~

One more thing

Finally, a warm message.

QuantumBit is dedicated to connecting humans and AI, and is also willing to help you connect with jobs. If you have any AI companies or positions that interest you, we can also help with "internal recommendations". Welcome to contact us - qgod001 , the quantum assistant .

Well, there is no agency fee.

Reference link:
https://www.reddit.com/r/MachineLearning/comments/enul63/d_how_does_your_company_interview_machine/