By Harish Doddi, CEO, Datatron
Artificial intelligence for business is changing the way we do work, yet companies are slow to adopt enterprise AI. A recent MIT study found that just 11% of executives plan to integrate machine learning (ML) capabilities into their products or services within three years. For many executives, ML seems complicated – like something they want to adopt but that seems too hard to actually pull off. Machine Learning Operations (MLOps) offers a way for companies to improve their chances of success with ML. Essentially, it’s a way of using models to systematize the ML lifecycle by defining processes to make ML development more reliable and productive. However, that doesn’t mean it’s always smooth sailing. There are a few key things CEOs need to understand about MLOps to make sure it succeeds for their organization.
Why a company might struggle with MLOps
MLOps requires dedication; you need to dedicate people and resources to such an undertaking. You need a dedicated team, and that team needs to grow over time. Sometimes, enterprise leaders treat this process much too casually and assume it will all work out, but it doesn’t.
Along with this, you need to invest in people with the right skill sets. It’s not just a matter of throwing bodies at the problem. MLOps is a skill that needs to be developed because to really do analog, you need to understand the machine learning portion of the AI models, but you also need to understand the operations portion. It can be difficult to find people who understand all of this. More often, you’ll find someone who is more operations-focused, and they don’t understand how ML works, so there’s a knowledge gap in general. It is harder to get skilled people who understand both aspects, but as a leader, you can balance bringing in new talent with developing talent from within.
Successful deployment of MLOps also requires a certain amount of planning ahead, just like most new technologies do. You need to carefully consider the various contingencies and possible outcomes before you initiate deployment; to ensure your organization is prepared in advance.
Adjusting culture for MLOps success
One of the most important adjustments to an organization’s culture when introducing MLOps is being able to demonstrate the separation of duties. Generally, when machine learning is involved, there are many different people are involved. Traditionally, AI was seen as a project for the AI group or the data science group. But these days, that’s not the case.
For instance, an engineer is interested because MLOps is involved in production environments. So, you’ll need to make sure you can carefully separate duties – because in some cases, the priorities for the data scientists aren’t the same priorities for the business leaders. And those might not be the same priorities from an operations standpoint.
Data scientists want to get the best possible outcome for the model decision. But from an operations perspective, they are all about achieving reliability and stability for the model. For example, if the model doesn’t respond within a certain timeframe, what are the consequences of that, or if the model is behaving in a strange manner? Unlike human decisions, when models make decisions, they can have drastic impacts very quickly because of the scale at which they’re operating.
Traditionally, AI has been viewed as being the purview of a siloed set of people, and it’s only they who are the stakeholders. But these days, that’s not the case. It’s not only data scientists or modelers. There are many stakeholders involved: operations, engineering, line of business – even the regulatory compliance people. So that’s why, rather than having a one-size-fits-all approach, each of them will have different priorities.
The priority for a data scientists might be very different from a buyer or an operations person. It can be very different to a businessperson. How do you bridge this gap is a key question because everybody has a different view of how things need to be done. In reality, in practical applications, everybody needs to work together. That’s the culture adjustment to aim for first.
What CEOs need to understand about MLOps
From the outset, CEOs need to understand that models are an iterative process. When someone develops the first version of the model, it’s not the final version; it’s the first draft. When stakeholders push it to production, and they see how it behaves, they learn from that and then take what they learn back to the development environment. They tweak it and make the second version. It’s a highly iterative process. Maybe one model is experiencing 99% accuracy, but under what conditions and under what assumptions?
That’s key to think about, because in a production environment, many things are changing, including data and behaviors of users. So, the things they observe in the development, they may not observe in production. They may learn something new; they may unlock some new insight. So, it’s always an iterative process.
Second, you need to adopt MLOps processes. This journey is important, and it’s key for the success of AI, because things are going to be more difficult the longer it takes to adopt these best practices. Here’s one example: If data scientists have the right set of tools in their development environment, they can move quickly. They can iterate fast on their models, but they don’t see the same thing in the production environment once people get involved. So, this behavior unnecessarily creates friction between the data science teams and other parts of optimization – engineering, operations, infrastructure and other parts of the organization. That is why the faster you can adopt best practices and have standardization, the better off you’ll be in terms of easing friction that could occur down the line.
Third, auditing is happening across huge volumes of data and across different business units, and it can happen in terms of models, too. You need to be able to show evidence and accountability for any questions the auditing team might ask. For instance, if the model loses money during a particular time period, you explain why that happened, and what – if any – actions were taken.
Start now
Research suggests that enterprise AI and ML adoption are moving at a glacial pace, despite all the hype around these technologies. While many AI and ML deployments fail, in most cases, it’s less of a problem with the actual technology and more about the environment around it. Successful ML adoption requires the right skills, resources, and systems, and using MLOps can give your organization a considerable competitive advantage here. Use the recommendations above to make the necessary culture shifts and remember the iterative nature of this journey. Thus equipped, you’ll be well on your way to enjoying the benefits of MLOps.
Harish Doddi
About the author
Over the past decade, Harish has focused on AI and data science. Before Datatron, he worked on the surge pricing model for Lyft, the backend for Snapchat Stories, the photo storage platform for Twitter, and designing and developing human workflow components for Oracle.
Harish completed his master's degree in computer science at Stanford, where he focused on systems and databases.
The views and opinions expressed herein are the views and opinions of the author and do not necessarily reflect those of Nasdaq, Inc.