Solutions

Productionized Machine Learning Creates a Competitive Advantage

Prioritize better systems over better models.

The commoditization of model algorithms means models alone are unlikely to provide competitive differentiation. Bespoke, ad-hoc models serve as a proof-of-concept for the potential of AI/ML initiatives, but an artisanal approach to deployment and iteration will not generate sufficient material value for the enterprise.

Today, companies that operationalize their machine learning achieve the most success. Uber has over 1,000 models in production concurrently, enabled by the development of their own pipeline and workflow management platform, Michelangelo. Booking.com has over 150 models in production, while CapitalOne conducts over 80,000 big data experiments a year. Facebook developed Pytorch as an AI framework and runs an average of 4,000 models daily, with 1,700 in production and 93% of new models developed with the framework.^x,xi,xii

Enterprises don’t have to reach the level of these digital titans to benefit from MLOps. MLOps focuses on culture, design, and infrastructure to enable the rapid development, deployment, and iteration of high-quality machine learning models while still allowing the freedom to explore and experiment. Data science teams that implement MLOps development practices into the machine learning pipeline may realize such advantages as:

Faster, more focused model improvement
More time for data scientists to experimentally improve models
Increased model throughput
More resilient, adaptable systems
Faster time to deployment
Quicker response to errors
Increased user satisfaction
Reduced risk of project failure
Extended life of prediction service
Reduced likelihood of catastrophe
Reduced development and operational cost

While many enterprises have embraced the machine learning value proposition, few companies currently generate value from their data science initiatives.

Manage the Machine Learning Lifecycle with MLOps

Experience the power of operationalized machine learning.

Despite some bleak statistics and anecdotes, the future of machine learning is bright. After all, machine learning is a nascent field, and best practices are typically hypothetical and experimental in nature. Software development faced a similar predicament 20 years ago. Development cycles experienced long feedback loops that caused obsolescence of requirements by the time of deployment. In the last decade, the software development community has adopted the DevOps philosophy to operationalize enterprise app creation.^xiii

The DevOps practice unites development for software systems with their operation using a culture of best practices and iterative improvements through continuous integration and continuous delivery. Enterprises use DevOps principles to develop and deploy roust, reliable software in shorter cycles.^xiv

Machine learning systems are themselves a type of software. As such, pragmatic leaders should take inspiration from the DevOps philosophy for their machine learning development strategy. They should implement similar practices adapted to the idiosyncrasies of machine learning models.

With the effective design and implementation of MLOps principles woven into each aspect of the machine learning development lifecycle, leaders can achieve the scale and reliability required for successful software development systems.^xv

Figure 3 MLOps systematizes deployment while giving data scientists the freedom to experiment. Image: Neal Analytics