What are we talking about when we talk about MLOps?
Nvidia is a company best known for its graphics processing units (GPUs). Also, given GPUs' ability to train deep learning models, Nvidia naturally became a benchmark in artificial intelligence.
Suppose you consider "ensure" to be an ambitious word. In that case, you may be reassured by the fact that Nvidia, and not just your humble server, defines MLOps as a set of acceptable practices for running AI successfully.
When we delve into more specific topics, it is always helpful to comment on what concepts are supposed to be familiar to the reader. The idea is that you can join us regardless of what you know about artificial intelligence or data science, but if you're feeling a bit lost going forward. I suggest you start here.
To understand MLOps, it is essential to start with the context in which this concept emerges. Some people comment that the magnitude of the solution has to be proportional to the size of the problem, and, as we will see below, the problem was not minor. This required, therefore, a solution with high standards.
So what was this context?
To avoid the typical introduction of "data is the oil of the 21st century", I prefer to tell you the historical progression of artificial intelligence and machine learning in the company. Below we can see a timeline of what happened in broad strokes in the industry (and I say broadly because, strictly speaking, we know that the concept of artificial intelligence emerged in the 50s, before cloud providers).
As we can see in the graph, before the popularity of cloud computing, there were already areas dedicated to analyzing data on a massive scale. However, the cloud's emergence allowed companies of different sizes to access high processing capacities without the need to invest considerable amounts in infrastructure.
This "democratization" of computing capacity opened the door for analytical areas to explore artificial intelligence projects. Companies knew that significant competitive advantages were now at hand, just a data scientist away ... or not?
Well, in reality, it was not as simple as expected. When trying new processes and technologies, there is always a period of adaptation, and for the industry, the figures were overwhelming:
According to a survey conducted by NewVantage of analytical benchmarks in the industry, it is shown that 77% of them confirm that the business adoption of AI and Big Data initiatives is still a challenge.
According to VentureBeat, 87% of data science projects never reach the productive phase of development.
Gartner forecasts that by 2022 only 20% of analytical insights will have a business actionable.
The exposed statistics are just a sample of the difficulties companies have had to obtain the expected value of artificial intelligence. This is why we call this adaptation process, "the fall".
If we compare the problems in software development and the number of years it took to start developing good practices and appropriate frameworks associated with these projects, it really seems natural to have gone through this stage.
However, given the vast global adoption of analytics and machine learning areas, learning on the go quite quickly. And it is in this context of accelerated learning that MLOps arises.
What is MLOps?
While we do not detail all the components involved in putting an artificial intelligence project into production or automating, the MLOps set of practices explains this on their own. In the graph, we can see a condensation of these principles and practices:
Below I will list each of them:
Reproducible and versioned: This point is crucial since it is not only necessary to version code, as, in conventional software development, we must also version models and data. Besides, given that all of these components can change, it is key to generate the notion of versions of each of these components (like Git for code)
Auditable and interpretable: Having reproducible and versioned components allows us to demonstrate which components of the data we used to assure auditors that the pertinent legislation is being complied with.
Packaging and validation: Given the number of existing frameworks, it is vital to package versions in each environment. With that, we will ensure that the code works beyond the local environment of each data scientist. It is also necessary to check the latency of the services. The model in production must meet the requirements that the project implies.
Deployment and monitoring: Being able to put the models into production is a tremendous challenge per se, so the implementation must consider the principles discussed above. It is also necessary to monitor the models' behavior because their quality tends to decline over time.
What do we learn?
It is interesting to analyze as a whole each one of the elements necessary to have successful developments in this field. Since this analysis shows that there were already practices in the development of conventional software replicated (such as versioning), others had to develop on the fly (such as versioning is of models and data).
Because it is a relatively new concept within the industry (being that AI is already quite new), there are many other ways to approach MLOps. Different approaches include reviewing it by layers (data, models, development, and operations) and can be approached from the specific roles in an analytical team and even from the frameworks.
But as academic texts say, when something is important, but it is not the focus to detail it, we will say that it is out of scope. But so that you do not think that it is lazy on the part of the writer, we will leave you the recording of the webinar in which we talk about MLOps and review all those details in a greater degree of depth.
It's in Spanish, but don't worry. You can activate the CC option marking the auto-translate, and you are good to go.
If you got this far, we think you might be interested in understanding what Imagemaker services are about, so I leave you our communication channels: our website, Instagram, LinkedIn, and of course! Follow us at Medium.
Machine Learning Consultant
WEBINAR: Ensure the Success of your AI Projects