Author: Melvyn Yap, PhD, Senior Machine Learning Researcher at Max Kelsen
The Elusive Research Path
Completing a PhD (Doctor of Philosophy) requires an incredible amount of focus, persistence, personal sacrifices, and quite often, a healthy dose of pure luck. It is common belief that earning the highest possible level of academic qualification will surely and automatically translate to securing a research job, and that your life is guaranteed to be stable and financially rewarding. But is it?
The brutal reality is that the vast majority of PhD graduates do not end up continuing a career in research, let alone becoming an independent researcher. This is quite contradictory to the common belief that the sole purpose of completing a PhD is just that — entering a life dedicated to pursuing the inquisitive realms of research. Yet, the number of PhD completions in Australia has grown from 3,933 in 2001 to 9,064 in 2017 (1). If their futures are so statistically grim, then why do so many students still choose to pursue PhDs? One can reasonably assume the massive enrolment is due to a host of external factors, such as the undeniable effectiveness of university campaigns, the relative ease of obtaining government funded scholarships, the irresistible enticement by prospective PhD supervisors, and quite simply, because it is the ‘next natural step’ after a Masters degree.
Committing several years of your life solely to completing a PhD is not a trivial matter, even if the motivations, as mentioned above, can be. Those qualified into PhD programs tend to be some of the brightest students in the cohort. As such, I strongly believe that the real reason for this commitment is none other than simply wanting to do research — to follow their curiosity of exploring the unknown, to have the ability to put wild hypotheses to the test, and/or to contribute to the wealth of human knowledge. To achieve these goals, and to be sufficiently recognised as a capable researcher, the time-tested PhD program remains as the only known path. Evidently, Albert Einstein, Stephen Hawking, Marie Curie, Alan Turing and other great scientists of our time have all completed a PhD prior to achieving their greatest work. This deep sense of fulfilment and willing tradeoff of so many years is a true testament to how important research is to individuals such as you and I.
Venturing into the unknown — the alternate path to pure research
This article is intended primarily for current PhD students wishing to continue a research career after completing their PhD, but can also be extended to current postdocs whose thoughts are wavering at their unknown future in academic research. While current statistics tell us that there are more PhD holders planning on switching their current research roles to non-research ones than vice versa (2), suggesting a poor perceived job security, there is no real reason for it to remain this way. In recent years, there has been a steady rise of science funding from outside of academic institutions (3), with a sizable proportion coming from the industry. Research work performed outside of academic institutions will be referred to from here onwards as ‘industry research’. Here, I present my perspective as a recently transitioned researcher from academia to industry. Specifically, from an academic neuroscientist to a machine learning (ML) researcher. As you are reading this, keep in mind that the experience described here is in the Australian setting, which may or may not apply to your locality. From someone who has walked the same path, felt the same apprehension and has held tightly to the same dreams as yourself — this is my story.
Misconceptions of working an ‘industry’ job
I am a Senior Machine Learning Researcher at Max Kelsen. During my PhD days, many conversations were had with my peers about what’s next after our PhD. An academic postdoc position is what immediately crosses everyone’s mind. When industry research is mentioned, there is a general, accepted perception of what it entails. Alarmingly, most of which, as I later discovered, are blatant misconceptions, such as:
- You need a business degree. None of my team members in the research group at Max Kelsen has an actual business degree. Particularly within the artificial intelligence (AI) industry, you will find that the most common degrees possessed by an ML researcher, ML engineer, or data scientist are one of the science, technology, engineering and mathematics (STEM) degrees, not business. Diversity of knowledge is more valued than any form of conformity in qualifications.
- You need a computer science degree. While having formal training in computer science would certainly be advantageous, the successful completion of a PhD serves as a solid testament to one’s ability to rapidly and independently acquire new skills and knowledge. With a bit of determination, one can easily upskill whilst on the job, using some of the many online offerings such as massive open online courses (MOOC) at a tiny fraction of the cost of university degrees.
- Overqualified individuals or PhDs cost more to hire. This might have been the case in the past, but employers have become increasingly aware of the value that PhDs can bring into their workforce. Unlike universities, there are no distinct pay grades within most commercial companies. Instead, pay is more appropriately scaled to the value you practically or perceptually bring to the company (how critical or relevant your current domain knowledge or possessed skillsets are). This is often estimated by means of performance during your interview, probationary period, ongoing employment, and even your ability to negotiate your desired salary. This latter point, in some cases, is a significant plus. This is because you are given greater control over your salary than what you would observe in fixed grading of universities.
- There is no autonomy. In an academic research role, you typically retain a lot of freedom to explore interesting, self-driven topics that satisfy your innate curiosity. This boundless freedom is not necessarily lost in the industry setting. In fact, I would argue that industry researchers are given greater creative, scientific licence, as industry funding comes with fewer constraints and greater variety of what is seen as successful deliverables (e.g., white paper or blog publications, and not just overly tedious peer-reviewed manuscripts).
- There are no more nerdy, science talks. Science talk occurs daily in both formal and casual settings in the life of an academic researcher. As it turns out, the same also holds true while working in a research team within a commercial setting. In fact, people working in the AI industry generally spend more time engaging in science and tech conversations than any other topics out there!
- You have to adhere to strict business dress codes. Business attire is usually a routine requirement for large, traditional corporates, or a general expectation for meeting business clients. But in a research role within an industry setting, you could simply show up to work in t-shirts and shorts and nobody will bat an eye — your attire is fortunately not a whole representation of your professionalism.
The wonders and woes of working an industry job
So, what else is different about working in the industry as opposed to academia? Here is my honest review of working as an ML researcher in the industry.
- Work-life balance. I listed this item first as I felt that this is by far the biggest change of working in the industry versus academia. My working hours are highly respected here at Max Kelsen. Never have I been told to work longer hours than needed. This is in stark contrast to academia, where spending long hours at work is not only deemed as necessary, but those who spend the most hours at work are often even applauded.
- Roles are more clearly defined and respected. In a typical academic research group, team members usually comprise researchers (postdocs, postgrad students, interns) and only one or two administrative staff. Oftentimes, the researchers are required to run various administrative tasks at the expense of lost time focused on science work. In a commercial company, there are usually multiple teams, each handling a different aspect of the project. Depending on its size, a larger project could involve data engineers, ML engineers, developers, researchers, project managers and marketers, with each team member working in the space of their defined role. Each area of the project is looked after by experts in the field, and every role is respected for their unique value and contribution.
- Working on diverse projects. As part of a research group not confined within a domain-specific faculty or school means that the areas we get to work in can be excitingly diverse. For example, some of the current and upcoming projects at Max Kelsen involve studying the survivability of corals from the Great Barrier Reef, predicting the outcomes of immunotherapy treatment, optimising quantum computers, and characterising the neural responses to environmental stimuli. There are no rules to which projects you may work on, and providing you show interest, willingness and aptitude, the possibilities are seemingly endless.
- Working in a multidisciplinary team. As mentioned, the privilege of working on diverse projects comes from the team being equally diverse in educational backgrounds. The research team at Max Kelsen currently consists of researchers trained in quantum physics, mathematics, electrical engineering, neuroscience, and genetics. There is no one field that is more indulged in than the other. This diversity is a true testimony to frequently used, but not often practically delivered, interdisciplinary research.
- The satisfaction of seeing project outcomes deployed in the real world. While the typical end goal of most academic research projects is to be published in reputable journals, most of the projects I am currently working on are centred towards having our research insights and ML models used in the real world, generating actionable results and helping people lead better lives. Thus, although we still strive to disseminate our scientific achievements, we have a greater variety of available communication media, such as this blog, and greatly value our practical outcomes.
- Grant-writing and other fund-seeking activities are still required. Research is costly. An important consideration when seeking an industry research job is to find out where the funding for your projects will come from, as this may have an impact on job stability. You also need to do due diligence understanding the company and its history. There are commercial companies that have the capacity to completely fund their internal research teams (think of big techs like Google or Facebook), companies that rely entirely on external research grants (no different from university research), and others that can flexibly switch between the two to ensure consistent research funding (Max Kelsen falls into this last category). The sustainability of your role can depend on funding, thus finding out ahead is crucial. Great news is that more likely than not, most of these are offered as permanent positions. As a consequence, so long as the company is thriving, your position should be secure (no more 12-month contracts, should you choose so!).
- Be prepared to work with people with no research experience. Unlike academia, not everyone you work with in an industry job has been exposed to research work. So be prepared to sometimes take on the role as a research mentor and perform quality checks by allocating time to review the work of others. At times you may find yourself clashing with your colleagues as a result of this lack of understanding the complexities of research.
- Working on diverse projects (if you prefer specialising in one domain). Whilst working on diverse projects can be exciting as mentioned in the ‘wonders’ section above, some of you may prefer to only focus on one area. Therefore, if there is an expectation from management to be involved in multiple projects across multiple domains, you may end up feeling some inner resistance, and subsequently, dissatisfaction towards your role.
- The frustration of gaining access to data. It can be challenging to convince data providers that the data requested will not be used for commercial purposes when you are working under the banner of a commercial company. It is possible to circumnavigate this blocker by establishing a formal collaboration with an academic institution, but you will have to be prepared for these kinds of push backs during a project.
What does it take to work in the industry as an ML researcher?
At this point, it is important to be reminded that the transition I personally made was two-fold: 1) from academia to industry, and 2) from neuroscience statistical methods to ML. What I have described thus far reflects my experience of the former, while the following section focuses more on my experience moving from a non-ML background to ML.
Although it may not be immediately obvious, PhD graduates have a mass array of useful skills that are highly sought after in industry jobs, such as in the field of AI/ML. Some of the most valuable skills include critical thinking, scientific writing, hypothesis formation, statistical testing, data visualisation, and presentation skills. With this strong portfolio, there are only a few more additional skill sets that are really needed prior to starting work as an ML researcher. These can be broadly grouped into hard skills or soft skills.
Soft skills or traits
- The motivation for rapid and constant learning. If you are genuinely passionate about ML, your capacity for rapid learning will naturally expand. There are plentiful online courses, code libraries and methodological articles to guide your progress and keep you growing as fast and as far as your heart desires.
- An appetite to solve complex problems. Increasingly more complex problems are exploring the use of AI to find solutions. In fact, many of the world’s most complex problems are already, or beginning to, rely on the capabilities of AI for an answer. Therefore, there is no escape from encountering complex problems as an ML researcher.
Hard skills by importance
Coding skills. ML researchers are heavily involved in coding. Be it reviewing the codes of others, implementing new algorithms, automating code runs, fixing coding bugs, conducting exploratory data analysis, or merely generating plots for reporting purposes.
- Python. If you are like me and had some experience with coding during your PhD, you are likely trained to use Matlab and Matlab only. Due to the proprietary nature of Matlab however, the industry has taken a hard steer away from this programming environment. Instead, open source languages, such as Python is the language of choice for most AI and software companies, Max Kelsen included. I would highly recommend picking your favourite beginner’s MOOC and start learning Python. The Python official user documentation offers the best coverage of Python knowledge. NB. It is worth noting that a variety of ML projects directly related to fields, such as biology, human health, or genomics will still take advantage of highly developed capabilities of languages, such as R. So if your academic career was centred around it, nothing has gone to waste.
- Github version control. Working in the industry means frequently working as part of a team. When multiple coders are involved in a single project, multiple versions of codes get produced, so effectively managing them is crucial. Github is by far the preferred choice for version control. Once again, there are many MOOC offerings on git version control so I would suggest checking this out.
- Command line. Terminal (macOS and Linux) is a command line system that quickly runs commands that control your operating system. It is simple to learn, yet incredibly powerful. Similarly, there are many tutorials online that provide easy to follow lessons for mastering the command line.
Machine learning skills. Unlike the field of software engineering and web development, machine learning has not yet reached the same maturity level curriculum-wise. As such, it can be hard to filter through the hundreds of seemingly well-developed lesson plans to identify which one is best to follow. Here are a couple of general guidelines that can help you to navigate through the plethora of resources:
- Pick one highly rated MOOC. My personal recommendation would be the practical deep learning course offered by fast.ai. As of today, this course is free and the learner should aim to complete the whole course.
- Read quality ML books. Published books are often of much better quality than the average MOOC. My personal recommendations for building ML knowledge are the Deep Learning book and Elements of Statistical Learning.
- Project-based learning. Completing an ML project takes you on a journey of the whole ML pipeline, from data cleaning to model training and testing to interpreting the results, which directly reflects the processes of a real world ML project. Data science competitions such as Kaggle presents a good opportunity to apply your knowledge and build your skills by completing a real world ML project.
Cloud computing skills. Private companies are increasingly migrating from physical servers to the cloud, while academic institutions still rely largely on physical servers. The biggest cloud service providers are currently Amazon Web Services (AWS), Microsoft Azure, and Google Cloud. It would be highly advantageous to first get familiar with managing cloud resources, and could be decisive when seeking positions outside academia. The best possible evidence and credential for you to acquire is a cloud certification (e.g. AWS Certifications, Google Cloud Certifications).
Completing a PhD is a fantastic achievement, no matter what. The journey alone teaches many valuable skills, many of which are readily transferable to non-academic settings and across multiple domains. In this article, I have provided the perspective of a former academic researcher on some of the long-held misconceptions of working in the industry, the wonders and woes of working in the industry, the additional skills required for a seamless transition, as well as recommendations on how to obtain them to make the move from academia to industry a smoother journey. It is in my view that a research career in the AI industry undoubtedly aligns with many PhD graduates’ goals. Especially if you want to continuously challenge yourself with impactful scientific problems without the constraints and pressures deeply affiliated with a more traditional academic position. If you have to take anything out of this article, is that there is more than one research path after a PhD, and I am a living and thriving testament to this.
If you would like to know more about what it is like working in this field, please connect with me on LinkedIn — let’s start a conversation.
- Australian Mathematical Sciences Institute & CSIRO Data61’s Ribit.net (2019). Advancing Australia’s Knowledge Economy: Who are the top PhD employers? http://bit.ly/2GUhBX1
- European Science Foundation — Science Connect (2017). Project report: 2017 career tracking survey of doctorate holders. https://www.esf.org/fileadmin/user_upload/esf/F-FINAL-Career_Tracking_Survey_2017__Project_Report.pdf
- Mervis, Jeffrey (2017). Data check: U.S. Government share of basic research funding falls below 50%. Science. doi:10.1126/science.aal0890