How To Destroy the DeepMind Research Quiz
DeepMind is a bleeding-edge UK Artificial Intelligence (AI) company that aims to solve Artificial General Intelligence (AGI). They became a household name when their Go-playing program AlphaGo beat Lee Sedol, the reigning world champion.
The Quiz
Note: I passed the research quiz in June 2019. Their interview content could have changed between now and then (I doubt it though).
The research engineering position at DeepMind is highly technical. To ensure that candidates have a strong technical background they administer a two-hour-long quiz.
The quiz consists of four parts: Computer Science, Maths, Statistics, and Machine Learning. With such a broad scope for questions, it requires an almost encyclopaedic amount of knowledge.
Preparing for the quiz seemed to be overwhelming. It didn’t help that the list of preparation materials they give out consists of two textbooks, three online courses, and links to lecture notes for The University of Cambridge first year maths and statistics courses.
Reading two thick textbooks is already unrealistic. In the section below I advise you on what you realistically need to do to destroy this quiz.
Preparation
After submitting my Bachelor’s thesis, I had one month to prepare for the quiz. To figure out what I had to study for the most efficient preparation, I solicited advice from people that worked at DeepMind and people that had done the quiz before.
The majority of my preparation centred around the Deep Learning Book by Goodfellow et al. I also watched David Silver’s Reinforcement Learning (RL) course but, this was not necessary as the RL section of the quiz was optional.
I created an Anki deck to help me memorize everything I needed for the interview. Creating and memorizing this deck gave me a deep appreciation and understanding of the fundamentals of Deep Learning.
To get a copy of the DeepMind Anki deck download it from my Google Drive.
Computer Science
For computer science questions, I would recommend going over the standard content from a Computer Science degree.
This part of the quiz required knowledge of the standard computer science degree topics such as threading, data structures, deadlocks, sorting algorithms, programming paradigms and networking. Again, you only need to know how to define things. So, there is no need to read the entirety of CLRS.
Maths
The mathematics questions had a heavy focus on concepts from linear algebra, calculus, and numerical methods. You also have to differentiate or integrate an equation by hand in this round. So, remember to brush up on the chain rule and integration by parts.
For a recap on algebra, calculus rules, trig identities, I would recommend reading and answering the questions of the math notes from Lamar University. The notes are comprehensive and well-produced. They make it easy to quickly learn the concepts and put them into practice.
To develop excellent intuition for linear algebra concepts, I highly recommend watching the Essence of Linear Algebra by 3Blue1Brown. Having a solid foundation in linear algebra will make learning maths for Machine Learning concepts a lot easier.
Statistics and Machine Learning
I spoke to a former DeepMind Research Engineer and they told me to read the Deep Learning book.
Following his advice, I read the book and added a card to the deck for anything that could be phrased as a “what is” question.
The questions for this section ranged from defining a specific matrix and its properties to explaining how different architectures work.
Conclusion
In the month that I studied for the quiz, I learned a lot. I was delighted when I passed the quiz. But it felt way better knowing that I developed a whole new level of understanding for my field.
Unfortunately, I didn’t make it past the last interview (pre-onsite) but I had a great experience during all of the interviews.
I hope that my Anki deck or the linked content in this article helps you succeed with DeepMind’s interview process. Or, any other interview process for Machine Learning based roles that you go for in the future.
Omar Reid — You can find me on LinkedIn and Twitter.
Reddit Resources
Past Questions from Reddit user
General Discussion on Reddit
