One of the most important things that you can do to have an impactful career is to develop a clear vision of your long term goals. This is difficult in machine learning because your work needs to strike a balance between being achievable and novel. Furthermore, it needs to be motivating to you personally, and impactful to others.

Choosing who to work with

One of the most important steps in your career is evaluating the visions of different potential mentors/managers and ultimately deciding who to work with. First, you want to look for people who work on problems that you find to be motivating and important. Second, there needs to be room for growth within that field. In order to build your ability to choose better collaborators over time, it is beneficial to understand how your mentors built their career over time. What motivated them to do each of their past projects? How did they scope out intermediate milestones to pursue their larger goals? What problems do they consider to be not worthwhile (sometimes what you choose not to work on is as important as what you choose to work on)? While you can’t just copy the career trajectory of your mentor, you’ll be better positioned to carve out your own path if you understand why they did what they did.

Maintaining a consistent direction with increasing impact over time

Since people are looking for the best person to solve a particular job or investigate a certain area, it pays to be excellent in one area rather than to be good in a number of areas. You should aim to be an expert in a particular ML subfield (e.g. language translation, robotics) or a ML technology that solves a specific business problem (e.g. recommendation algorithms). Furthermore, a successful career is composed of a series of milestones that move towards a specific direction with increasing scope over time. In reality, to get your foot in the door, you may need to do projects with goals that aren’t particularly motivating to you. Regardless, you will get the opportunity to demonstrate competency and an ability to grow. If you complete smaller projects well, people will be willing to entrust you with larger, more important projects in the future. Additionally, you should aim to do projects in the same area with increasing scope and impact. However, in reality, not every project you do will be along one career trajectory. In this case, you can still make progress by reflecting on what you found most motivating in the past and try to make better decisions in the future in terms of what to work on.

Achievable yet novel

A central challenge of developing a career trajectory is to do work that is both achievable and novel. One key is to identify an area where the research ideas and technology have only recently gotten powerful enough to make meaningful progress. For instance, the breakthrough in image classification came when innovations in neural network optimization came together with sufficient computational power (e.g. train a ~10 layer neural network with ~100 channels per layer on a million images for a number of epochs). In order to explore this idea further, I’ll consider two researchers (now assistant professors at top universities) who started their Ph.D’s around the time that I did. For reference, a watershed moment in deep learning was late 2012 when Krizhevsky et al. developed the first convolutional neural network approach to outperform classical methods on image compression.

Chelsea Finn at Stanford has become an expert in developing intelligent robots. Back in 2014, there were exciting successes in classification, but robotics had seen relatively few applications of deep learning. Robotics is difficult because you need an actual robot and the advances in neural networks (algorithms and accelerators) were relatively new. Thus the research area was prime for growth due to the convergence of different ideas and technology. Her work has investigated algorithms that can enable an embodied agent to learn its sensor-motor relationships through exploration. Furthermore, there is the promise of more robust industrial robots.

Song Han at MIT has become an expert in the topic of efficient deployment of neural networks. Back in 2014, again, neural networks were achieving impressive research results, but there were many unanswered questions on the topic of practical deployment. Song and collaborators were able to identify that the size of the parameter files was a large problem and doubled down on developing network compression methods. Their work made it much easier, for instance, to add a neural network to an app that needs to be downloaded onto many devices with limited disk space. Since then, he has been effective in working with industry partners to identify other deployment challenges related to speed and different hardware.

Final Thoughts

All that being said, identifying an area with the potential for a lot of growth is only one step of the process. These researchers were successful because they were additionally able to scope out specific milestones within their domain and find actual solutions to those intermediate goals. If you liked this article and want to learn more about how to identify the best milestones and create solutions to those problems, connect with me on LinkedIn to see my future blog posts. If you want to get personalized feedback on how to make better decisions at all stages of your career, check out my coaching page.