AI is changing the world. But how quickly can the average data science team deploy ML models? Depending on who you ask, this might be a small handful or even just a couple of models a year.
According to a new report from Deloitte on the tech trends of 2021, MLOps, or the application of engineering discipline to automate ML development, is set to significantly shorten the time it currently takes to put a newly developed ML model into production.
As organizations industrialize AI and shorten development life cycles, this shift could put an end to the current era of “artisanal AI” that relies on the efforts and talents of select individuals. Instead, the focus will shift towards engineered performance and a consistent approach through MLOps.
This development could lead to a golden age for AI. Deloitte cited research from Cognilytica to say that the MLOps market is expected to expand to nearly USD 4 billion by 2025.
A need for change
According to Deloitte, many organizations are hamstrung in their efforts by clunky development and deployment processes that stifle experimentation and hinder collaboration between stakeholders – from product teams, operational employees, to data scientists.
In one survey of nearly 750 business decision-makers, a mere 8% consider their companies’ ML programs sophisticated. The figures look dire among those surveyed. Forty-seven percent fail to take their AI initiatives out of the experimental phase, while more than one quarter (28%) of ML projects fail.
According to IDC, the primary reasons cited for failure range from lack of the necessary skill set, lack of production-ready data, and integrated development environments. Finally, one in five (18%) say it takes more than three months to bring an ML model to where it produces business value.
What is MLOps
It is worth noting that software development faced similar development and operational challenges two decades ago. This led to the birth of DevOps, which entails the standardization and automation of application development, deployment, and management of software. This resulted in dramatically improved quality and greater development efficiency.
MLOps borrows from principles of DevOps to automate the development and deployment of applications to AI models. It brings together crucial components such as data management, automated model development, retraining, code generation, and continuous development and monitoring.
Deloitte says MLOps can encourage experimentation and rapid delivery, helping enterprises to industrialize machine learning. For instance, new techniques, supported by better data organization can reduce the process of customizing and adjusting the way models learn (model tuning) to days or even hours.
MLOps should not be confused with AIOps, or Artificial Intelligence for IT Operations, though. The latter is concerned with the application of AI to automate the management of IT and accelerate digital transformation. You can read more about AIOps from Gartner here.
The future is bright
Don’t expect MLOps to be the magic bullet to accelerate ML into the stratosphere, however. MLOps differ from DevOps in significant areas and has its own challenges that must be addressed. For a start, ML practitioners must deal with complex, data-related issues not typically faced in software development, such as accountability and transparency, compliance, and AI ethics.
Finally, AI bias is an area that is garnering attention with the growing use of AI. Fortunately, this is yet another area where MLOps procedures can make a difference. As noted by the report: “Without MLOps procedures in place, it would be infeasible, if not impossible, to prove proper data handling or use in response to an external inquiry.”
Much work remains to be done, but the future is bright. As summed up by Swami Sivasubramanian, the vice president of machine learning at AWS: “We are entering the golden age of machine learning, with adoption increasing across all customer segments.”
Check out the full Deloitte report here.
Image credit: iStockphoto/Alexyz3d