Seventy-four percent of respondents to a recent O’Reilly Media survey said they consider machine learning (ML) and artificial intelligence (AI) to be a game changer, indicating it has the potential to transform their job and industry. Indeed, 61% report ML/AI as their company’s most significant data initiative for the upcoming year. While reports show that many organizations struggle to get beyond test phases (53% in a recent Gartner report), those that successfully move AI/ML into production reap significant benefit. Enter MLOps – or Machine Learning Ops – the practice of successfully moving ML models into production.
MLOps in a nutshell
It is generally said that machine learning is the inverse of programming – the process of hard coding logic into a software program. For example, let's say we write a software program to calculate client credit scores. This is typically done by examining client data attributes, evaluating the attributes, assigning more weight to some attributes and less weight to others. From here, we would calculate a score. While this logic is generally given to us by domain experts or business analysts, we would gradually fine tune the logic, based on how our scoring performs in the context of busines outcomes.
Conversely, ML logic is not prescribed and hard coded, but rather derived from the data – and the more data, the better. Let’s continue with our credit scoring example. To achieve the same results, we would look at volumes of client data along with their credit scores to derive a function that predicts credit scores. (An ML model is a function that is mathematically derived from the data.)
Given the input and the output, we can approximate a function. And, once we have an approximate function in hand, we can then begin the process of tuning the weighting. We want to tune it in such a way that the error rate between the function’s range of output generated to that of the output values in the training dataset is minimal. Once the tuning is completed, the model should be validated with tests and deployed. In this way, the program that generated credit scores has now been reverse engineered using the data it generated. The original program itself is a function or function of functions. While there are many ML techniques, this is an example of supervised learning.
Why do I need MLOps?
As you could guess by the “Ops” nomenclature, MLOps takes a page from DevOps, the popular set of practices that help organizations achieve faster time to market. As you likely know, DevOps automates manual processes and reduces waste in the value stream, enabling teams to release features faster to customers. When valuable units of work are delivered in small time-boxed sprints, feedback loops are faster, in turn, helping businesses adapt quickly to changing market needs.
These DevOps principles apply to deploying ML models, too. MLOps creates an infrastructure that enables the automated building, training, evaluating, deploying and monitoring of ML models. These steps align with DevOps’ continuous integration, continuous testing, continuous deployment and continuous monitoring processes, offsetting for the fact that the way code is built, tested and released is different from how ML models are built, tested and shipped.
As a result, MLOps helps organizations undertake and execute more ML projects and deploy models faster to production. MLOps infrastructure and automation helps teams release new versions faster and more frequently. MLOps helps models adapt to feedback with agility, with changes to models driven by feedback from monitoring, changes to data and even changes to the model itself. Ultimately, MLOps reduces the cost of new ML projects and decreases maintenance costs for existing productionized models.
Clearly, ML is a boon to helping businesses reach key objectives. From extracting better quality information to reducing costs and mining more value from existing data sources, ML can help businesses grow the topline and reduce the bottom line. For example, as early as 2016, Netflix reported saving $1 billion from its ML algorithm that recommends personalized content to subscribers.
MLOps helps organizations achieve these goals across deployment and automation, model reproducibility, scalability, security, greater team collaboration, and ongoing monitoring and management. Said another way, MLOps can help streamline the ML lifecycle just as DevOps streamlines the software development lifecycle.
Because IT leaders often find it challenging to scale ML projects and achieve this level of business success, in our follow up article on MLOps, we’ll discuss the MLOps lifecycle of a ML model, MLOps phases and components and we’ll share a real-world implementation example. Stay tuned – or subscribe to our newsletter below -- to ensure you don’t miss our discussion of how to facilitate greater performance, scalability, and reliability with MLOps.
Post Date: 08/09/2021