Greg Benson has been a University of San Francisco professor teaching computer science for the last 20 years. He’s also part of a quiet effort to solve one of the field’s growing problems.
During semesters, he spends one to two days a week at SnapLogic, a cloud integration company, and works full-time there during school breaks. Each year, he offers 10 of his machine learning master’s students the chance to intern at SnapLogic and work on AI research projects. If they do well, they can end up with a job. About a third of SnapLogic’s engineering team are former interns. “This has been wildly successful for us,” Benson said.
This is a useful recruiting method when even the tech giants are struggling for talent. But there’s also another innovation afoot: SnapLogic’s deal with Benson also shows how companies can support academics and make sure there are still professors to teach the next generation of AI researchers.
The problem is stark: Almost 60% of U.S. computer science PhDs take industry jobs after graduation, up from 38% over the last decade, according to data from the National Science Foundation published by the Wall Street Journal.
Another study out of Worcester Polytechnic Institute found that in many universities, there were more faculty positions open for AI and related areas than were filled. AI and associated fields was one of the tracks with the highest discrepancy between PhDs produced and positions filled, with a net of 6% more positions filled than PhDs produced. The problem is further compounded by the fact that it takes three to five years to earn a doctorate in computer science.
It’s unclear how much talent is even out there to begin with. The Montreal-based startup Element AI published a study that found 22,000 AI researchers with relevant PhDs since 2015. That number goes up to 90,000 if you count degrees awarded pre-2015, but the field is changing so fast, recent grads are often thought to have the most relevant research. Tencent, the Chinese internet giant, published its own estimate that put the figure between 200,000 to 300,000 people able to contribute to AI research.
The hunt for top talent has become a hiring arms race that uses everything from offers of seven-figure deals to Flo Rida performances at industry events. Even researchers at AI nonprofits backed by big money are making more than a million dollars a year. The talent is highly concentrated. Six technology companies—Google, Microsoft, NVIDIA, IBM, Intel, and Samsung—employ over half of all deep learning specialists, according to a December 2016 report by KPMG.
Benson said the University of San Francisco has struggled to hire enough professors to teach AI. Being in the Bay Area, the combination of expensive home prices and the strong draw from industry is hard for practitioners to ignore when looking for work. The University of San Francisco gets 700 to 800 applications a year to the school’s machine learning master’s program, and takes less than 10%. They could potentially accept more students or expand offerings, but they don’t have the faculty to do it.
Some of the biggest names in AI maintain a heavy load of teaching and industry work, or cycle between the two. Geoffrey Hinton joined Google and splits his time there with the University of Toronto. Yann LeCun divides his time with New York University and Facebook. Andrew Ng balanced being a professor at Stanford University and leading Baidu’s AI group before leaving the latter to start his own fund and other projects.
Carnegie Mellon University’s Andrew Moore has jumped between industry and academia throughout his career, but recently announced he is leaving CMU to lead Google Cloud. That position is open because Fei-Fei Li recently left the position to go back full-time to Stanford. (Li, who was on leave from Stanford for nearly two years, had always planned to return, a Google spokesman told Business Insider in June.)
Before leaving Carnegie Mellon, Moore tried to fight back against the siren song of industry by making it easier for faculty to go back and forth between the two worlds. Moore told the Wall Street Journal that he estimates that 10% to 20% of faculty will take leaves of absence to go to industry or even found a startup.
This might not be a scalable solution. Top researchers in the right cities might have the resources that allow them to work intense schedules—childcare, assistants, important meetings planned around their time—but those perks aren’t available for everyone.
Sofus Macskassy, VP of data science at HackerRank, a technical recruiting platform, says that balance is nearly impossible. He speaks from personal and professional experience—he previously taught at the University of Southern California while working for a startup in Los Angeles and hired AI talent at Facebook.
“Realistically speaking, you don’t have time to do both jobs really well,” he said.
To create highly trained AI researchers, it’s not just teaching students. It is also advising and helping research get published. That type of work is much different than the applied AI most workplaces are interested in.
“It’s hard to straddle both,” he said.
Balancing both also means making sure to sort out the issues of intellectual property (IP) and patents, and what is developed where. It can be quite difficult to navigate IP legalities as the creator is stuck between two legal entities that often both want the IP rights. For the most part, these legalities often get resolved by the lawyers for the university and company without the professor having to get involved. It gets trickier if the professor is starting a company. Macskassy said this depends on the university’s organization that manages IP. If that department is less experienced, it can become a headache. But it is a process that has plenty of precedents to rely upon.
Macskassy said the talent shortage was the bigger conundrum. If left unaddressed, the U.S. will run into shortages that puts it at risk of losing ground in innovation and research to other countries like China.
“Companies are shooting themselves in the foot in the long term,” Macskassy said.
Another solution might be to support academia with more research grants. With the Trump administration’s cut to funding, that might be a pipe dream for the time being. But where the U.S. federal government is failing, some tech giants are trying to staunch the flow by directly funding university departments and teaching students.
Facebook’s AI Research program, called FAIR, is working with academics that let them split their time between Facebook and their home university. The program recently expanded, with nearly two dozen researchers now maintaining a dual affiliation; how the researchers split their time is up to them. IBM launched a program called Cognitive Horizons Network that partnered with six schools to work with researchers and students while keeping academics in their university positions. The project saw 70 peer-reviewed AI publications in its first year.
Last month, Microsoft unveiled its own effort to tackle the shortage of AI-related skills in business and academia. Professor Chris Bishop, director of Microsoft’s Research Lab in Cambridge in the U.K., told ZDNet that the launch of two new training programs is a “multimillion-pound program in which we’ll invest in PhD scholarship, postdoc, internship, and consultancy positions.”
Of course, all this focus on PhDs and master’s degrees might be a losing strategy, no matter what compromises are made. Not every problem needs a fancy algorithm, and understanding the goals of a certain business comes from experience, not a degree. One group making this argument is Fast.ai, a nonprofit that offers free AI courses while putting out original research has the goal to make deep learning more accessible, or, as their cheeky slogan says, “Making neural nets uncool again.” A group of students from Fast.ai beat teams is made up of researchers from Intel, Google, and other tech giants in a challenge to train the fastest and cheapest object recognition algorithm.
“They are fighting over the same few people where they could be looking at a much broader group,” said Rachel Thomas, a cofounder of Fast.ai, who herself switched over to machine learning from finance after taking online courses.
Thomas also points out that the talent shortage is partially a perception problem that ignores the fact that training employees in AI could be accomplished with a variety of online courses like Fast.ai or Coursera (which, incidentally, is a company cofounded by Ng while he was at Stanford; it now hosts his machine learning course). These already-in-place workers have a deep insight into the company, the data available, and the problems the company wants to solve with AI. Not all of those issues need cutting-edge resources, especially with all the open source tools and AI software currently out there.
“Companies think, ‘Oh, I have to hire the Stanford PhD,’ and that’s not actually what they need,” Thomas said. “In-house talent is being undervalued right now.”
Benson agrees that companies aren’t tackling the problem as head-on as they could. Taking on researchers with fewer degrees, providing continuing education, and training internally are all company-based solutions that need more resources. There is only so much that the university system can do before companies need to pull some weight.
“Academia has responded, but industry doesn’t know what it’s going to take,” Benson said.
This story was updated to clarify that Facebook has more than two dozen reachers with dual affiliations, not 12.