Last year, the fastest-growing job title in the world was that of the machine learning (ML) engineer, and this looks set to continue for the foreseeable future. According to Indeed, the average base salary of an ML engineer in the US is $146,085, and the number of machine learning engineer openings grew by 344% between 2015 and 2018. Machine learning engineers dominate the job postings around artificial intelligence (A.I.), with 94% of job advertisements that contain AI or ML terminology targeting machine learning engineers specifically.

This demonstrates that organizations understand how profound an effect machine learning promises to have on businesses and society. AI and ML are predicted to drive a “Fourth Industrial Revolution” that will see vast improvements in global productivity and open up new avenues for innovation; by 2030, it’s predicted that the global economy will be $15.7 trillion richer solely because of developments from these technologies.

The scale of demand for machine learning engineers is also unsurprising given how complex the role is. The goal of machine learning engineers is to deploy and manage machine learning models that process and learn from the patterns and structures in vast quantities of data, into applications running in production, to unlock real business value while ensuring compliance with corporate governance standards.

To do this, machine learning engineers have to sit at the intersection of three complex disciplines. The first discipline is data science, which is where the theoretical models that inform machine learning are created; the second discipline is DevOps, which focuses on the infrastructure and processes for scaling the operationalization of applications; and the third is software engineering, which is needed to make scalable and reliable code to run machine learning programs. 

It’s the fact that machine learning engineers have to be at ease in the language of data science, software engineering, and DevOps that makes them so scarce—and their value to organizations so great. A machine learning engineer has to have a deep skill-set; they must know multiple programming languages, have a very strong grasp of mathematics, and be able to understand and apply theoretical topics in computer science and statistics. They have to be comfortable with taking state-of-the-art models, which may only work in a specialized environment, and converting them into robust and scalable systems that are fit for a business environment. 

As a burgeoning occupation, the role of a machine learning engineer is constantly evolving. The tools and capabilities that these engineers have in 2020 are radically different from those they had available in 2015, and this is set to continue evolve as the specialism matures. One of the best ways to understand what the role of a machine learning engineer means to an organization is to look at the challenges they face in practice, and how they evolve over time. 

Four major challenges that every machine learning engineer has to deal with are data provenance, good data, reproducibility, and model monitoring.

Challenge 1: Data Provenance

Across a model’s development and deployment lifecycle, there’s interaction between a variety of systems and teams. This results in a highly complex chain of data from a variety of sources. At the same time, there is a greater demand than ever for data to be audited, and there to be a clear lineage of its organizational uses. This is increasingly a priority for regulators, with financial regulators now demanding that all machine learning data be stored for seven years for auditing purposes.

This does not only make the data and metadata used in models more complex, but it also makes the interactions between the constituent pieces of data far more complex. This means machine learning engineers need to put the right infrastructure in place to ensure the right data and metadata is accessible, all while making sure it is properly organized. 

Challenge 2: Good Data 

In 2016, it was estimated that the US alone lost $3.1 trillion to “bad data”—data that’s improperly formatted, duplicated, or incomplete. People and businesses across all sectors lose time and money because of this, but in a job that requires building and running accurate models reliant on input data, these issues can seriously jeopardize projects. 

IBM estimates that around 80 percent of a data scientist’s time is spent finding, cleaning up, and organizing the data they put into their models. Over time, however, increasingly sophisticated error and anomaly detection programs will likely be used to comb through datasets and screen out information that is incomplete or inaccurate. 

This means that, as time goes on and machine learning capabilities continue to develop, we’ll see machine learning engineers have more tools in their belt to clean up the information their programs use, and thus be able to focus more time spent on putting together ML programs themselves.

Challenge 3: Reproducibility

Reproducibility is often defined as the ability to be able to keep a snapshot of the state of a specific machine learning model, and being able to reproduce the same experiment with the exact same results regardless of the time and location. This involves a great level of complexity, given that machine learning requires reproducibility of three components: 1) code, 2) artifacts, and 3) data. If one of these change, then the result will change. 

To add to this complexity, it’s also necessary to keep reproducibility of entire pipelines that may consist of two or more of these atomic steps, which introduces an exponential level of complexity. For machine learning, reproducibility is important because it lets engineers and data scientists know that the results of a model can be relied upon when they are deployed live, as they will be the same if they are run today as if they were run in two years.

Designing infrastructure for machine learning that is reproducible is a huge challenge. It will continue to be a thorn in the side of machine learning engineers for many years to come. One thing that may make this easier in coming years is the rise of universally accepted frameworks for machine learning test environments, which will provide a consistent barometer for engineers to measure their efforts against.

Challenge 4: Monitoring

It’s easy to forget that the lifecycle of a machine learning model only begins when it’s deployed to production. Consequently, a machine learning engineer not only needs to do the work of coding, testing, and deploying a model, but they’ll have to also develop the right tools to monitor it. 

The production environment of a model can often throw up scenarios the machine learning engineer didn’t anticipate when they were creating it. Without monitoring and intervention after deployment, it’s likely that a model can end up being rendered dysfunctional or produce skewed results by unexpected data. Without accurate monitoring, results can often slowly ‘drift’ away from what is expected due to input data becoming misaligned with the data a model was “trained” with, producing less and less effective or logical results. 

Adversarial attacks on models, often far more sophisticated than tweets and a chatbot, are of increasing concern, and it is clear that monitoring by machine learning engineers is needed to stop a model being rendered counterproductive by unexpected data. As more machine learning models are deployed, and as more economic output becomes dependent upon these models, this challenge is only going to grow in prominence for machine learning engineers going forward.

Machine Learning Engineer: An Evolving Role

One of the most exciting things about the role of the machine learning engineer is that it’s a job that’s still being defined, and still faces so many open problems. That means machine learning engineers get the thrill of working in a constantly changing field that deals with cutting-edge problems. 

Challenges such as data quality may be problems we can make major progress towards in the coming years. Other challenges, such monitoring, look set to become more pressing in the more immediate future. Given the constant flux of machine learning engineering as an occupation, it’s of little wonder that curiosity and an innovative mindset are essential qualities for this relatively new profession.

Alex Housley is CEO of Seldon.