What I learned from my Data Science Internship at Volvo Group
During the summer I was an intern in the Advanced Driver-Assistance System (ADAS) team at Volvo trucks. I worked on developing a prototype for an automatic verification system of their intelligent speed assist function. In other words, I worked on traffic sign detection and recognition with very little labeled data for the task at hand. Challenging yes! Also lots of fun. Here are the main lessons I take with me.
Big company = lots of complex processes
Volvo Group is a worldwide company and only the site in Gothenburg, where I worked, has a couple of thousand employees. What I found most wonderful about working at such a large company was the shared vision and sense of family among the people there. In corporate meetings, the leaders gave a sense of purpose and direction and made everyone feel needed. The scale does, however, come with some drawbacks. Working with programming I soon realized that nothing was customized to the teams’ individual needs rather than to satisfy everyone’s needs, systems were patched up and combined to fulfill sometimes relatively simple requests. And if a small change were to be made, in my experience, it generally had to be passed by some busy person in another team which at times caused some frustration. However, as a summer worker, I was also surprised by how accommodating everyone was whenever an issue occurred and that there was always someone to guide you through whatever was needed.
The Agile Mindset
Coming from studying four years of mathematics and computer science I have done a lot of projects, assignments and essays. Usually, the process is quite linear from start to finish with a short sprint toward the deadline. However, if you have worked at a company in some kind of software development you realize this is not the way most products are developed today. The two principles I’d like to take with me from the agile way of working to other parts of my life are:
- Iterate on it
It is better to have something working that is far from what you want than a few perfect details summing up to nothing at all. As soon as you have something that works, a minimum viable product, you can iterate on it. I take this with me to writing, studying, cleaning, really anything.
- Design an acceptance criteria
When are you done? I have lived my life letting deadlines say when I’m done and sometimes procrastinating school assignments so that I know I will not be able to spend too much time on them. What if you instead make a conscious decision to decide when you are done with projects, assignments etc? At Volvo, this is done with a short description of the tasks needed to be done or the result to be achieved to be able to move the task from In progress -> Done. A concrete example is that for this article I have written a short outline, I will write a draft in one session and then edit in a second session and then I am DONE.
KPIs — what gets measured, gets managed
Before my work at Volvo, the word KPI was not one in my vocabulary. So, I will define it for you: KPI = Key Performance Indicator. In other words, it is the metric you are interested in measuring for a certain function you are developing. One of the responsibilities of the team I worked with during the summer was to support other teams in developing KPIs to measure their functions’ performance. In machine learning I have always been aware that metric is important, however, my experience from the vehicle industry taught me that the metrics they optimized towards often are based on standards and regulations that need to be fulfilled rather than standard metrics to assess a deep learning model. One of the challenges in the project I worked on was therefore to implement a computable metric based on the regulation document for Intelligent Speed Assist systems in Europe.
Working with real data
Whenever I talk to someone who works in data science in the industry this comes up as one of the challenges in the transfer from studying to working. As students, we are used to a reasonably sized, relatively clean and labeled dataset. What I faced at Volvo was a gigantic and unlabelled dataset for a task that required supervised learning. The solution we came up with used external open-source data sources, some simulated data and real-world data for testing. With more time and resources a self-supervised approach would have been a really interesting path to take. I also found myself on the other side of the machine learning problem, data creation, spending an entire workweek labeling data to even have a test set.
Altogether, I take with me that working in the industry has a lot more to do with efficiently making use of data than designing complex model pipelines. Key lesson: If you are taking the path to be a data analyst/scientist in the industry be prepared to swap your expected percentages of time spent in Pandas vs PyTorch! We are truly in the generation of big data now and many companies, I would believe, have a lot more data than they can productively make use of today.
Wrap up
The points mentioned above do in no way compile a complete list of what I learned during the internship. Rather I gained many insights and improved my technical skills just from the experience of working intensely for two months on a real-world computer vision problem. This would, however, not have been the case if the team I worked with had not been so welcoming and shared their knowledge. So lastly I’d just like to extend a big thank you to these people that contributed to making my summer internship both fun and valuable!
