A machine learning learning PhD doesn’t only open up some of the highest-paying jobs around, it sets you up to have an outsized positive impact on the world. This comprehensive guide on machine learning PhDs from 80,000 Hours (YC S15) will help you get started.
80,000 Hours is a nonprofit that researches careers with social impact and provides online advice. The guide is based on discussion with six machine learning researchers including two at DeepMind, one at OpenAI, and one running a robotics start-up. Check out the highlights below.
Why do a machine learning PhD?
Machine learning involves giving software rules to learn from experience rather than directly programming the steps it takes. It’s a hot field, and some of the hype is true — powerful ML systems could radically transform society, for good or ill. We could see benefits such as more accurate medical diagnosis and autonomous cars that slash traffic deaths. But the negatives include the potential for greater inequality, mass unemployment and even catastrophic accidents.
With the machine learning skills you can gain in a PhD, you can help shape this powerful technology and apply it to important problems. You might work on:
• Social impact start-ups such as building alternative economic indicators for the developing world.
• Technical research such as building systems that learn their goals from human preferences, which could make them less likely to do dangerous things. Read more about this kind of technical research in our profile on research into risks from artificial intelligence.
• Policy research such as modeling AI arms races. Read more about this in our guide to working in AI policy and strategy.
If you want to focus more on applying machine learning rather than researching new techniques, however, then a Masters degree is often sufficient.
How do you get in?
You’ll need an undergraduate degree in a quantitative subject such as physics, maths, computer science, or engineering where you should have covered probability and statistics, multivariable calculus, and linear algebra. You’ll also need to know how to program, either from courses or teaching yourself.
Many people focus on grades, but showing that you can do good research is key to getting admitted, so dive into research as soon as possible. As an undergraduate, get summer research positions. If you do a master’s degree, do one with a strong research component and try to get publications before you finish.
It’s also critical to get good letters of recommendation, ideally from people who are known to the people assessing your application.
Although grades are less important you’ll still need a GPA higher than about 3.6 in the US and a 1st or high 2:1 in the UK. For applying to US universities you’ll need to get >= 95th percentile on the quantitative part of the GRE.
What topics should you study?
According to researchers at DeepMind and OpenAI, these techniques look promising for building powerful artificial intelligence systems:
• Deep learning: it’s been making progress on problems that have been difficult for previous machine learning methods
• Reinforcement learning: it’s a closer fit for humanlike trial-and-error learning than other machine learning methods
How can you test out this path?
Here are some ways to test your fit, roughly in order:
- Talk to people who are doing machine learning PhDs.
- Learn from books and courses such as those in this list.
- Read papers and implement models from them. If you find this addictive and you can implement models quickly, that’s a good sign. Check out important papers to do this with this in deep learning and reinforcement learning.
- Do summer research internships and possibly a master’s degree that includes research projects. Ask your research supervisors whether they think you’re a good fit for machine learning research.
Read the full guide for more detail and to get free one-on-one advice.