I recently underwent the entire process for DeepMind, so I thought I'd share here since there are very few resources on the web about their new process. I was rejected at the end of the process with the feedback that I didn't have enough experience in ML experimentation and modeling. Recruiter reached out to me on LinkedIn saying they have a position open in Mountain View for DeepMind for Google. Coding Screens: - Two rounds, Google style questions which I'm pretty sure were picked from Google's internal question bank. - First round: LC medium. Follow up was a LC medium/hard which I was not asked to code once I explained the solution. Then was asked to code a LC easy. Follow up was a LC easy/medium which I was again asked not to code once I explained the solution - Second round: LC Hard and the question is not on LC. I gave a factorial solution, improved it to an exponential solution with memoization. Later found that there is a polynomial solution, apparent to those who have a strong competitive programming experience. I don't think they expect the polynomial solution though. - Different from google as you have to run the code and it's expected that by the end you have a solution which runs in CoderPad. ML screens: - If one clears the coding rounds, they move forward to ML rounds - 2 ML design style rounds - First round probed the depth of ML knowledge. Started with questions around probability. Was asked a number of balls style question on bayes theorem. Proceeded to the design question. It's not enough to know things, you should know the mathematical intuition as well. E.g. it's not enough to say that you prefer L1 regularization over L2 since L1 results in sparse features, you need to give the mathematical intuition in terms of how gradients affect the weights in L1 to cause sparsity. - Second probed width of ML knowledge. Hardest ML round I ever gave. Another design style question, but they kept adding constraints once I gave a solution. In total, we discussed like 7-8 different techniques across the board. Over the course of the interview at least 7-8 constraints were added over every solution I gave. Virtual Onsite: - If you clear the ML rounds, they move forward to virtual onsite. - Two non technical interviews with TLs where they asked about my background and my experience with ML experimentation. I have some experience from Amazon in ML experimentation, so I told them whatever experience I had. Got the feeling during the interview they want someone with more modeling experience than mine. - One behavioral with their people & culture partner (fancier term for HRBP). Standard questions like why do you wanna join DeepMind. In total the process took about 6-7 weeks. I was able to expedite a little by telling them I have deadlines due to google and salesforce. Edit: Removed an earlier section ranting about the process. I thought that since I was forwarded to the onsite, my ML screen performance was strong. But based on some of the comments here, it doesn't look like that's the case. TC: 195K (L5) YOE: 4 + MS (2 YOE in US) Background: My experience at Amazon was in the intersection of ML and engineering with more bias towards engineering Edit:
This is pretty cool! You’ll definitely be better prepared one you start looking with all the practice you’ve gotten
Thanks for the kind words! The prep definitely helped me crack the Google interview, but unfortunately the team match is going nowhere there.
Do they move forward with you to virtual on sites because you have passed the coding and ml rounds or they just move forward no matter what the feedback is?
After both the screens they mentioned that they'll keep me updated after gathering feedback. I'm pretty sure they only passed me on to ML screens after coding rounds since they took about 2 weeks to get back to me after the coding screens. For the ML screen, since I was on a deadline they might have combined onsite feedback with the ML screen feedback, so I might not have done as well as I thought there.
I’m curious - what do you mean by constraints in the design rounds. Is it considering different models? Or adding things like dropout? Thanks !
Thank you for sharing. A few questions: 1. How did you prep for the ml interviews? 2. I don’t see any research engineer positions available, do they fill up fast or something? 3. What level did they interview you for?
1. I eventually failed their bar so I'm not sure if I should be giving ML interview prep tips. But in case you're still interested I brushed up mathematical concepts from Deep Learning ebook and revised concepts of NLP since that is what I work on. I also got a little lucky that one of the design questions asked was something that I had worked on previously. 2. I don't think DeepMind is the best at updating their career portal with available positions. Based on the research I did, most of their positions are either filled through referrals or through internal google candidates. 3. They didn't mention any level. I'm guessing that part would have happened if I cleared the interview.
Thanks so much for your responses. Congratulations on the Google offer! Is that for an ML position as well? Did you have similar interview experiences there as well? Hope you get a good team matched soon!
> Personally, I felt like they should have looked at my profile before the technical rounds to determine my profile wasn't what they were looking for. If it's anything like Google's pipeline then the rounds aren't a binary pass/fail. You might've had a "weak hire" in previous rounds and the hiring manager decided to proceed with TL interviews in the hopes that you can make up for it by showing stronger than expected background. If you mean that they should've had TL interviews first then that wouldn't make sense considering how expensive that is. TL's hour costs the company way too much to be wasting on filter interviews and that would be true for any tech company.
Totally possible that my feedback from ML screens was a "weak hire", at which point whatever you're saying makes complete sense
Thanks for sharing your experience btw, not much is known about it even internally.
Good luck next time. I also went through the similar process sometime back and have a good impression. Almost all my onsite interviewers were ex- Google brain.
Hey did you get in to Google brain?
@icauraus PMed you :)
Thanks for sharing!
They used to do a quiz earlier have they stopped that
Yup replaced it with 2 coding + 2 ML screens
For fwiw, quiz happens depending on whether we apply for Deepmind for Google research, or Deepmind for Google. I interviewed for the former, and had 2 hour quiz, 1ML, 1 coding and then similar on-site process.
mad.warrio How did you prepare for coding ? LeetCode ? Google tagged questions or something else ? I am unable to find any DeepMind tagged questions. Did you do Cracking the Code Interview or Elements of Programming Interviews ? Please advise.
Google tagged questions on LC. Almost everyone in DeepMind came from Google so they use the same question bank for interviews.
Thanks for sharing! Did they indicate whether you may be considered for a different team in the future? Seems a bit harsh for somebody to pass all technical stages and then be ejected completely because they didn't match with *one* TL.
India
Yesterday
670
Who are these retards asking for dictatorship in India?
India
Yesterday
601
Modi is a legend, will be remembered for centuries to come
Tech Industry
Yesterday
1813
Quitting this Slave life
World Conflicts
Yesterday
1301
Why people waste their time behind Israel Palestine issues?
Tech Industry
Yesterday
1440
The end of Backdoor Roth?!
Thanks for sharing. Are you an MLE or AS at Amazon?
MLE