Joshua Wallace

Yesterday, a graduate alumna of my department, Ai-Lei Sun, hosted a meeting with graduate students and postdocs in my department, talking about her transition from a postdoc to a nonacademic job. She was visiting campus for reunion weekend and was kind enough to offer to do this. As background for her, she graduated with her Ph.D. two years ago and has been pursuing postdoc work since then. She was sought out by the company Orbital Insight with whom she had had some contact in years past and decided to transition out of academia and work for them.

Orbital Insight

Just a brief word on the company itself, since I want to focus mostly on what was discussed at the meeting. The company though definitely merits a few words.

Orbital Insight does exactly what the company name suggests: gathers insights from orbit. It purchases images taken from satellites, balloons, UAV’s, etc. and then uses computer vision and machine learning to turn pixels into useful data and then sells information from the data to its clients. As an example, they can count the number of cars in parking lots and use that to measure how well retailers are doing. I think this is a fascinating way to mine data, partially because it’s exactly what I do in my research as an astronomer: extract useful data from a series of images.

What We Talked About

The meeting was very well attended: 7 grad students and 7 postdocs, which represents about a quarter of each of those segments of the department. A lot of substantive and even spirited conversation ensued.

Ai-Lei talked a lot about nonacademic career prep and job search. She said that an internship is probably the best way for a grad student to make the switch to a nonacademic job, because then you can show that your skills have already been successfully applied outside of academia. However, an internship is not a possibility for all industry-bound grad students. Luckily, there are other ways to demonstrate the applicability and usefulness of your skills outside of academia. Ai-Lei suggested that a well-maintained GitHub might be the best way to do this, especially if you write code “for fun” that works on public, nonacademic datasets. That is an incredible advantage of GitHub (and the internet in general), being able to fully showcase your work to anybody in the world with internet access. It is definitely a tool that should be leveraged.

One interesting thing that was brought up (primarily by the postdocs in attendance) is that the time is limited for people with Ph.D.’s not directly related to data science to be able to find data science jobs. The argument is that many data science, AI, machine learning, etc. programs are springing up around the country, and so in a few years there will be a sufficiently large number of people formally trained in data science that there won’t be room in the industry any more for people not formally trained as data scientists. I was very surprised to hear this and personally disagree. There certainly are and always will be jobs for which a “formal” data scientist—one who has been formally and deeply trained in data science, statistics, machine learning, etc. as a student—are the only ones qualified. However, many data science jobs don’t need as deep of an understanding of data science theory or algorithm development as a “formal” data scientist possesses, and so someone less formally trained as a data scientist but also possessing other useful skills should be able to find a job for which their unique skillset is well suited, even with a flood of “formal” data scientists coming into the job market. As a personal example: I work a lot with images from telescopes. I know what kind of things can go wrong when using a telescope and the impact that such things have on the images. For any sort of data that is derived from images (especially images that consist of just a few point sources, such as images tracking satellites), I already know what to look for and I’m ready to hit the ground running. It would probably be easier for me to learn the data science tools I need to know to use the data effectively than it will be for a “formal” data scientist to learn about how focus changes, clouds, cosmic rays, saturation, etc. affect the data.

Another point is that the term “data science” is not precisely defined and definitions can vary greatly between industries, employers, and individuals. All it takes is one Google search to find a great variety of definitions. Is data science a collection of skills? A way of thinking? Is it focused more on data wrangling, data mining, or data analysis? Is it the intersection of algorithm development and big data, or the intersection of scientific thinking and advanced computing? Is it a combination of some or all of these? As I’ve looked through data science job postings I’ve noticed jobs that, though having the same job title, are completely distinct in what they expect the successful applicant to do. There is a great diversity of data science jobs out there, and I believe there will always be data science jobs well suited for someone from a particular background even if they are not formally trained in data science.

Interest from postdocs

As a last point, I was surprised at the number of postdocs who attended this meeting. It was originally envisioned as just a grad student event, but several postdocs told Ai-Lei they were interested, so, about 24 hours before the event, the invitation was opened to postdocs as well. Even with the short notice, half the audience were postdocs. There are about the same number of postdocs as grad students in my department, so this was a good showing from them. What surprised me most, though, was which postdocs attended. There were several whom I thought were firmly set on the academic path: well-established independent scientists making grand contributions to the field, some of whom even have prized fellowship positions. I guess I figured anyone who had made it that far on the academic track were very happy and satisfied with astrophysics and didn’t have any interest in leaving. That is clearly not the case, though.