The Machine Learning Skills Issue

Machine Learning (ML) experts are difficult to find on the market and data scientists with a statistical background are among the highest paid professionals in software, according to StackOverflow. The “war for ML talent” has been driving some companies to recruit whole enterprises (Microsoft acquiring Maluuba) or R&D departments specialised in this field (Uber with Carnegie Mellon). According to an analysis conducted by Geodesic Capital, tech giants like Amazon, Google, and Apple employ nearly a quarter of all top-tier machine learning talent available.

Indeed, ML requires knowledge in so many disciplines, from mathematics, probability theory, and statistics, to computer programming and data engineering, that an expert is necessarily a rare gem. Not to mention the many existing sub-fields of specialisation, such as computer vision, Natural Language Processing, Reinforcement Learning, signal processing, etc. It has been said that the dream of a data analyst is to become a data scientist, and that a data scientist will aspire to become an ML expert.

So how does one get to be an ML expert? The profession is so new that candidates receive education and training predominantly outside universities. According to a Kaggle survey (Kaggle is an international network platform for ML experts), 30% of 16000 respondents studied ML in college, while 66% were self-taught; 50% of all respondents have used online courses in any case. The proliferation of specialised courses outside traditional computer science Masters or PhDs is a challenge of its own, as it can be difficult to choose from the extensive list of options offered by Massive Open Online Courses (MOOCs). The choice of programming language can be another issue: should one specialise in Python or R? Should one also study Lisp and Prolog? Will C and C++ be necessary outside ML hardware programming? These are examples of questions that are hard to answer before the Machine Learning recruit knows what his or her application domains will be.

In a similar fashion, there are so many specialised software packages, libraries and tools to choose from that it may be difficult to know ex ante whether to begin with Scikit-Learn, Numpy, Tensorflow, Coffee 2, or Karas, and the field of Machine Learning may give the impression of being somewhat fragmented.

Consider an ML professional who is past these early dilemmas and is now being given a first assignment in an industry context. The question remains: which models and algorithms should one select for the problem and the dataset at hand? All algorithms have different trade-offs: whether neural networks will be preferred for their high accuracy, or decision trees for their interpretability, will be context dependent. In many instances, the differences in performance level of ML algorithms will not be so obvious. Should the ML specialist choose logistic regression, k-nearest neighbours, random forests, Support Vector Machine, or Naive Bayes? Or a mix of those? At the end of the day, only experience can provide the answer.

This is a big challenge for ML specialists: ML skills are not sufficient in themselves, experience with industry data and domain expertise are both necessary for the success of a ML project. Real-world and corporate experience are vital for solving deployment issues, as a successfully trained and tested ML system still needs to be integrated into the company’s existing data architecture.

Business acumen should not be forgotten among the many skills that the ML expert must display. Understanding what is important for the business, and what makes good business sense, can help define the right questions, or choose the best ML approach. Sometimes resource efficiency is as important as model accuracy. How to communicate the findings to non-specialists is another non-trivial challenge and should be considered a skill on its own. There is a need in every business for “ML translators”, ie professionals having both the necessary domain expertise and the sufficient technical expertise to understand the work of ML experts and derive insight from it.

Publicly accessible Trello board indicating the different skills that RSSB’s Risk and Safety Intelligence team identified last year for developing its data science capability

How to design jobs that make the most of these new breeds of data experts is not an easy matter. There is already a high rate of dissatisfaction among ML experts currently employed, perhaps due to frustration with bad data, or with a perceived lack of clear questions, or lack of technical knowledge among collaborators. 

This may sound like a lot of challenges. From a business perspective, the fundamental issue remains how to secure access to professionals with ML and data science skills, who also have the relevant industry experience and domain knowledge. 

One way to secure access to ML experts is to build the necessary partnerships with academia and set up mutually profitable placement and knowledge-sharing programmes. For example, RSSB is currently supporting a few Centres for Doctoral Training (CDTs), co-funded by the Engineering and Physical Sciences Research Council (EPSRC). The goal of the CDTs is to train the future workforce (and research force) to tackle research and innovation challenges across the engineering and physical sciences landscape.

In the area of Machine Learning, we are supporting the Autonomous Intelligent Machines and Systems CDT, led by the University of Oxford, which provides PhD candidates a comprehensive view of autonomous intelligent systems, combining theoretical foundations, systems research, academic training and industry-initiated projects and thus mixing both practical and theoretical aspects of intelligent machines and systems.

Another way is to establish and maintain informal networks of experts and collaborate with relevant think-tanks, such as the Alan Turing Institute, who are able to provide access to a pool of talent through modular or virtual teams.

Last, but not least, upskilling; reflecting on the origin of ML experts (physics, computer science, statistics, bioinformatics, chemical engineering), the FT wrote that Machine Learning is “the first new discipline to demonstrate the importance of lifetime learning”. Statisticians can be retrained in computer programming and new ML packages, or data analysts familiar with industry data can be retrained in statistics and probability theory. But there are many more opportunities within companies to find the right candidates. A liberal arts background is also fine, as long as the candidate has passion for Machine Learning and the tenacity to upskill as appropriate. After all, as Andrew Ng (Adjunct Professor at Stanford University, co-founder of Coursera and founding lead of Google Brain) explains, as the field of ML advances, it is getting easier for the non-specialists to break in, due to increased automation, open source tools and user-friendly packages.  

Andrew Ng co-founded and led Google Brain and was a former VP & Chief Scientist at Baidu, building the company’s Artificial Intelligence Group into several thousand people. Ng is an Adjunct Professor at Stanford University. He is also an early pioneer in online learning — which led to the co-founding of Coursera and deeplearning.ai. 

All three avenues to boost ML capabilities are applicable and worth exploring by railway organisations. And the third one should not be underestimated: what if the upskilling of current staff was the greatest opportunity for the railway industry to grow its ML talent pool? Any organisation that doesn’t include data skills in its plans for the training and development of both analysts and wider staff, is at risk of being left behind in the ML race.