MLops: The Key to Pushing AI into the Mainstream

One of the main roadblocks preventing the enterprise from putting artificial intelligence (AI) into action is the transition from development and training to production environments. To gain real benefits from the technology, this must be done at the speed and scale of today’s business environment, which few organizations are capable of doing.

This is why the interest in merging AI with devops is gaining steam. Forward-leaning enterprises are trying to blend machine learning (ML) in particular with the traditional devops model, which creates an MLops process that streamlines and automates the way intelligent applications are developed and deployed and then updated on a continual basis to increase the value of its operations over time.

Problem Solver

According to data scientist Aymane Hachcham, MLops helps the enterprise deal with a number of significant issues when it comes to effectively building and managing intelligent applications. For one thing, the data sets used in the training phase are extremely large and are continuously expanding and changing. This requires constant monitoring, experimentation, adjustment and retraining of AI models, all of which becomes time-consuming and expensive under traditional, manually driven development and production models.

To effectively implement MLops, the enterprise will need to develop a number of core capabilities, such as full lifecycle tracking, metadata optimized for model training, hyperparameter logging and a solid AI infrastructure consisting not only of server, storage and networking solutions but software tools capable of rapid iteration of new machine learning models. And all of this will have to be designed around the two main forms of MLops: predictive, which attempts to chart future outcomes based on past data and prescriptive, which strives to make recommendations before decisions are made.

Mastering this discipline is the only plausible way for AI to trickle down from the Fortune 500 enterprise to the rest of the world, says Greenfield Partners’ Shay Grinfeld and Itay Inbar. The fact is, upwards of 90 % of ML projects fail under current development and deployment frameworks, which is simply not tenable for the vast majority of organizations. MLops provides a dramatically more efficient development pipeline that not only reduces the overall cost of the process but can turn failures into successes at a rapid pace. The end result is that the barriers to AI implementation drop to a level that is comfortable for the vast majority of enterprises, leading to widespread distribution and eventual integration into mainstream data operations.

Early Success

MLops is still an emerging field, so it may be tempting to write it off as just another techy buzzword, says business analytics and data science consultant Sibanjan Das. But its track-record so far has been pretty good, provided it is designed the right way and targeted at the proper goal: to maximize model performance and improve ROI. This requires careful coordination between the various components that create an MLops environment, such as the CI/CD pipeline itself, as well as model serving, version control and data monitoring. And don’t forget to build robust security and governance mechanisms to minimize the risk of the ML model’s activities and the chance of it being compromised.

Even though MLops is designed for automation and even autonomy, don’t overlook the human element as a key driver of successful outcomes. A recent report by Dataiku noted that over the past year, companies have come to the realization that they cannot scale AI without building diverse teams that can implement and benefit from the technology. MLops should be a critical component of this strategy because it supports diversification in the development, deployment and management of AI projects. And just judging by Gartner’s MLops framework, a broad set of skills will be required to ensure that outcomes provide top value to the enterprise business model.

Even the most advanced technology is of little value if it cannot successfully transition from the lab to the real world. AI is now at the point where it must begin making a valuable contribution to humanity or it will become the digital equivalent of the Edsel: flashy and full of gadgets but with little practical value.

MLops cannot guarantee success, of course, but it can lower the cost of experimentation and failure, while at the same time putting it in the hands of more people who can figure out for themselves how to use it.