The field of machine learning has progressed quickly in the last decade, and there’s a big reason why: both academics and industry groups are working on the problem, often in competition with one another. Much ink has been spilled over the fierce rivalry for talent between corporate and academic research labs, but I’d argue that the two need each other for the betterment of the industry as a whole.
Consider this: academics by their very nature are obsessed with the optimal -- or creating as close to perfect systems as possible. They’re motivated by experimentation in pursuit of these optimal systems (and publishing research to validate their claims).
To contrast, companies are more focused on building strong customer relationships, and developing products based on solving real customer problems. In many cases in the software development world, worse is better. Rather than pursue a theoretical experiment on an unlimited timeline, the “optimal solution” is often neither required nor desired. Academic research serves as a source of ideas, techniques, and inspiration -- rather than a manual for product development.
This, of course, is a vast oversimplification of the dynamic between corporate and academic researchers in machine learning. Here’s a quick look at where the industry is at today.
Machine Learning's Current Status: Talent Wars
Many universities find their departments under siege by corporate research labs, who hire away top talent and make it nearly impossible for academic institutions to match salaries. While universities can only train so many PhDs to keep up with demand, top companies go where the talent is located. For example, in 2015 Uber hired 40 researchers from Carnegie Mellon for its self-driving car unit, which led to intense debate over the nature of the “partnership” and accusations of “brain drain.”
In attempts to serve multiple masters, many companies offer academics co-employment or dual affiliation roles, where they share time between corporate and university interests. Perhaps the most prominent dual affiliation is Yann Lecun, who is chief AI scientist at Facebook, but still retains part-time affiliation with New York University. Some academics argue that this approach does not work, since corporate and academic interests are in fierce opposition with one another. According to critics, in one corner is the academic pursuit of curiosity, and in the other is the need for immediate, tangible business benefit.
Why Machine Learning Needs Both Sides
In defense of corporate labs, some researchers say that the burden of chasing grants that plagues academia is lifted from corporate researchers, paired with an unlimited supply of data and computing power from which to work. While these benefits are clear, many academics counter that corporations could achieve similar gains by funding academic research grants (theoretically reducing the need to chase dollars), rather than taking talent away from universities.
As with other industries that are counterbalanced with academic and corporate interests (the pharmaceutical field, in particular, comes to mind) both academic and corporate labs are needed to push the machine learning industry forward.
Coming from a company founded by academics, I can say that practical applications of research are needed to fuel innovation for the industry. For example, if companies can ship products that are broadly adopted, the reach of academic research extends far further than its theoretical origins (academia gets off the island). Startup companies that are founded by and heavily employ academics apply a rigor to their approach that mirrors academia -- counter-balanced with the need to find practical applications. Schooled (literally) in the rigors of peer review and tenure competition, it’s not always easy for academics to shed the instinct to pursue the optimal and switch gears to shipping a minimum viable product.
In addition, there needs to be a two-way between larger corporate labs and academia in order for the industry to progress. On one hand, companies can provide contributions back to academia (in the form of data, access to computing resources, open source software, or endowments). On the other hand, academics can continue to pursue the optimal with unbound intellectual curiosity, providing their research for companies to apply to commercialized products.
It is possible to have both academia and corporations chase innovation in the same field without said field imploding on itself. My hope is that some harmony is found in the- battle for talent, and that intense career demand encourages even more students to pursue the field of machine learning.