With the rapid influx of AI and machine learning across the enterprise these days, teams looking to implement new tools and systems don’t necessarily know where to start. MLOps While practice provides guidance, it is important to know how to establish these efforts effectively.
As interest in AI and ML models has increased over the past year, many organizations want to: Scale your AI and ML efforts. MLOps is the practice of machine learning and DevOps. Similar to the role of DevOps for traditional software, it provides a framework for building and deploying ML models that aims to streamline model development and make it more reliable and efficient.
Yaron Haviv, co-founder and CTO of MLOps platform Iguazio, and Noah Gift, founder of AI education provider Pragmatic AI Labs, highlight this growing need for MLOps in their book. Implementing MLOps within the enterprise: A production-first mindset. Haviv and Gift provide detailed best practices and tutorials on MLOps initiatives and explain how to navigate them. We describe the key processes and highlight which aspects of the framework are most beneficial to those looking to integrate MLOps practices into their workflows.
“There are some basic components [of MLOps] Just as you can’t install a new appliance in your kitchen without a solid foundation in your home, adopting basic MLOps practices is the best way to ensure successful adoption of your model within your enterprise. It’s a method.
3 MLOps Best Practices to Consider
Haviv and Gift suggest some comprehensive tips to assist those looking to incorporate MLOps practices into their organizations.
1. Beyond model training
The MLOps pipeline consists of four main stages: data collection and preparation. Model development and training. Deploying ML services. and continuous feedback and monitoring.
As organizations begin to get serious about MLOps, model training is often the first focus. However, while training a model is important, prioritizing that process over other stages that are equally necessary for the model to work can quickly lead to problems.
“If you’re training a model and you don’t know how to reproduce what you just did, you’ve introduced a lot of nondeterministic behavior,” Gift says. “So there’s a risk in using this [model you created]Because what if you’re making important decisions based on this model?”
Having a clear methodology in place across all stages of MLOps helps teams avoid potential risks. With the MLOps framework in place, organizations can ensure that they are not implementing models into their workflows without first understanding and evaluating all core components.
2. Have a production-first mindset
A successful MLOps pipeline begins with the end in mind.in Implementing MLOps within the enterpriseHaviv and Gift describe this approach as a “production-first mindset.”
“A production-first mindset means having a critical thinking style when producing something,” says Gift.
Adopting a production-first mindset means teams constantly critically evaluate the entire MLOps process, from goal setting to deployment to ongoing monitoring. Maintaining a skeptical outlook and constantly monitoring the big picture can help model developers avoid tunnel vision and instead consider the pipeline holistically.
3. The less you can do, the better.
Another core tenet of a successful MLOps initiative is to reduce complexity as much as possible.
“The big piece is reducing the complexity of what organizations are doing,” Gift said. “The less we do for ourselves, the less complex the organization becomes. [and] Your organization is more likely to succeed. ”
Although organizations can reduce the complexity of their MLOps initiatives in a variety of ways, two practical options are to integrate continuous integration and continuous delivery (CI/CD) Build a pipeline and use a pre-trained model.
At the heart of CI/CD is a feedback loop that can inform the MLOps team of areas for improvement, Gift said. That feedback comes from a series of continuous improvement mechanisms such as automated linting, testing, and deployment.
“The CI/CD pipeline is like a truth-telling pill: It constantly examines and improves your code,” Gift says.
Another way to reduce complexity is to use pre-trained models instead of building ML models from scratch. With rapid technological advances in AI and ML, model equivalence is beginning to become a reality, Gift said. This means that many of today’s pre-trained models operate at similar performance levels.
Building your own models often doesn’t yield results. profitable ROI This is especially true for generative AI models that require large amounts of data and compute, which are time and resource intensive. Employing a well-researched, pre-trained model suited to your organization’s specific needs saves you time and money, allowing your team to focus on other aspects of your MLOps pipeline and leverage your organization’s strengths even more. You will be able to use it effectively.
When selecting among pre-trained models, organizations can consider standard characteristics such as functionality (for example, whether the deployment mechanism fits the intended use case) and price. Another important factor is potential negative externalities. Gift describes this as a drawback of integrating certain technologies into an organization.
This may include evaluating how ML models may negatively impact creativity and productivity. But it also ethical concernslike hallucination Prejudice, and increasingly concerning legal issues.Many AI and ML developers Litigation over copyright issuesunderstanding the legal implications of using certain pre-trained models is essential for organizations looking to implement those models.
These negative externalities can enter an organization as unquantifiable risks, Gift said. “If you use a model that contains copyrighted data and it turns out that the company that trained that model has lost the rights; [copyright] If I am suing, am I also infringing on this copyrighted data?”
It is essential to consider these questions when weighing the characteristics of one model against another. Look for models of organizations that are elevating responsible AI from the model development process to board-level executives.
In an excerpt from the book’s first chapter, Haviv and Gift explain how an effective MLOps strategy needs to be comprehensive and systematic.
MLOps: What is it? Why do we need it?
At the root of inefficient systems are interconnected bad decisions that worsen over time. It’s tempting to look for a silver bullet to solve an underperforming system, but that strategy rarely, if ever, pays off. Let’s consider the human body. Although there are many quick-acting treatments available to maintain good health, the solution to a long and healthy life requires a systematic approach.
Similarly, there is no shortage of advice on “get rich quick.” Again, the data contradicts what we want to hear.in don’t trust your intuition (HarperCollins, 2022), Seth Stevens Davidowitz shows that 84% of the top 0.1% of earners receive at least some money from owning a business. Additionally, the average age of founders is around 42 years old, and some of the most successful companies include real estate companies and car dealerships. These are not get-rich-quick businesses, but businesses that require significant skills, expertise, and wisdom gained through life experience.
Cities are another example of a complex system for which there is no silver bullet. WalletHub has created a list of America’s best-run cities, boasting a beautiful climate, home to the world’s top technology companies, and a 2022-2023 budget of $14 billion for a population of 842,000. . The budget is the same as that of the entire country of Panama, which has a population of 4.4 million people. As the example of San Francisco shows, profits and natural beauty alone are not enough to run a successful city. Comprehensive planning is required. Execution and strategy are key. There is no single solution that will make or break a city. WalletHub’s research shows a wide range of criteria for well-run cities, including infrastructure, economy, safety, health, education, and financial stability.
Similarly, for MLOps, it’s tempting to look for a single answer to getting your model into production, perhaps by getting better data or using a specific deep learning framework. . Instead, as in these other areas, it is essential to develop a comprehensive, evidence-based strategy.
What are MLOps?
At the heart of MLOps is continuous improvement of all business activities. In the Japanese automobile industry, this concept is called kaizen, which literally means “improvement.” When building production machine learning systems, this manifests itself both in the noticeable aspects of improved model accuracy and in the overall ecosystem that supports the models.
One good example of a non-trivial component of a machine learning system is business requirements. If a company needs an accurate model to predict how much inventory to store in a warehouse, but a data science team is creating a computer vision system to track the inventory already in the warehouse, it might be wrong. The problem will be resolved. No matter how accurate an inventory tracking computer vision system is, businesses have different requirements, and as a result, the system fails to meet the organization’s goals.
So what is MLOps? Combining machine learning (ML) and operations (Ops), MLOps is a technology that enables the efficient deployment of ML models in production environments to continuously improve business operations. Processes and practices to build, enable, and support. Like DevOps, MLOps is based on automation, agility, and collaboration to improve quality. If you’re thinking about continuous integration/continuous delivery (CI/CD), you’re right. MLOps supports CI/CD. According to Gartner, “MLOps is intended to standardize the deployment and management of ML models in parallel with the operationalization of ML pipelines: the release, activation, monitoring, performance tracking, and management of ML artifacts. Supports reuse, maintenance, and governance.”
Olivia Wisbey is an associate site editor at TechTarget Enterprise AI. She graduated from Colgate University with a BA in English Literature and Political Science and served as a peer writing consultant at the university’s Writing and Speaking Center.