Asking Machine Learning/AI hires to have fancy degrees is outdated. Here’s why.
Should your machine learning hire have a PhD? Do you need a PhD to work in ML?
I see PhD and Masters degrees listed as requirements in ML job descriptions all the time. The very first Google Jobs result I opened for “Machine Learning Engineer” required:
Ph.D. in Data Science, Machine Learning, Statistics, Operations Research or related field
M.S. in related field with 5+ years experience applying data science techniques to real business problems
Having a PhD or even Master’s degree is a really unusual ask for software engineers. We don’t expect this of developers working on networking or security or systems architecture or app development. So what makes machine learning so special?
Some might say that ML is uniquely complex and math-y, the domain of scientists rather than software developer hacks (ouch!).
I don’t buy it. ML may be tricky, but so is cryptography and distributed systems and graphics and tons of other topics in Computer Science. Yet we don’t require developers have PhDs to work on those things. I think it’s something else:
We often forget that machine learning is really new for most of us in tech. Just five years ago, my colleagues were still talking about deep learning as a dubious bet. Tools were difficult to use. Beefy hardware designed for ML was hard to get. Model quality was far from where it is today.
As a result, most of us didn’t learn about ML in school. Five years ago, Princeton, my alma matter, offered only around ~3 ML/AI classes. Machine learning was definitely not a “standard” part of your typical Computer Science curriculum, and you could easily graduate without learning much about it. Online resources were scant. Then, when AI suddenly became the new hotness, lots of folks rushed back into Master’s programs to fill the newly-relevant gap in their education.
With a shortage of ML talent, it’s no wonder that if you did want to hire someone with experience in the field, that person would probably be an academic.
Meanwhile, engineers began learning ML on the job. Even at Google, one of the world’s largest employers of PhD AI researchers, tons of engineers who work on ML products have limited prior experience with the technology. They learn through online resources, internal courses (like Google’s Machine Learning Crash Course), or by taking on small chunks of projects and learning as they go.
Five years is an eon in tech, and the data science landscape has changed. It’s much easier to learn machine learning outside of the classroom today than it used to be, and our toolset has become significantly more user-friendly (see PyTorch, TensorFlow 2.0, Keras). Paired with an enormous and growing ecosystem of online resources, the determined developer can give herself a hefty ML education without ever spending a dime.
The eligible Data Scientist/Machine Learning engineer hiring pool has changed, too. In 2019, the data science competition site Kaggle surveyed ~4000 data scientists. They found that while 52% of respondents had Master’s degrees, only 19% had PhDs. Meanwhile, the majority of respondents had only 3-5 years of experience, and skewed young (between 25 and 29 years old). There is a sizable and growing chunk of ambitious, self-taught data scientists just entering the job market.
We take for granted the fact that self-taught software engineers can be very talented (a recruiter for Google recently told me she interviewed a high schooler). Tech recruiters learned long ago that if they only hired MIT grads, they’d hire no one at all. So now we need to make a similar perspective shift in the way we hire data scientists.
This also means we need to start evaluating ML job applicants the way we evaluate software engineers. Instead of focusing on credentials, we need to spend more time building effective interviews that allow candidates to show off their skills no matter how or where they learned them. Building new criteria to hire ML engineers won’t be an easy task, but for employers, the payoff will be worth it.