Industry executives and experts share their predictions for 2019.  Read them in this 11th annual VMblog.com series exclusive.

Contributed by Don Roedner, Head of Marketing, Alegion

2019 Predictions for Machine Learning Projects

With all of the hype surrounding AI you might reasonably expect us to exit 2019 with streets full of driverless vehicles and call centers populated with chatbots. But reality, at least for machine learning projects involving computer vision, natural language processing and other use cases that rely on unstructured data, will be quite different. It's still early days for projects like these, and we're confident that 2019 will look more like this:

Most machine learning projects kicked off in 2019 will be the company's first

Most of our clients are in the Fortune 500. None of them has an ML system in production. Every project we are involved in is the organization's first. Executive- and board-level mandates will continue to drive the creation of internal AI labs, with lots of hiring and experimentation.

Data Scientist-driven pilots will deliver encouraging results, leading to pressure - and budgets - to deliver production systems

In our experience, newly hired or newly assembled data science teams first respond to their mandate by looking for big model-driven improvement steps. They use open source data that's available on the internet, they use pre-trained models, and they can pretty quickly get the underlying confidence in the model up to a certain point, without having to do more specific training.

Take a use case like frictionless shopping. Retailers can quickly say they've already demonstrated that they can track 70% or 80% of the actions going on in the store.

At this point they've gotten far enough to demonstrate that there is an opportunity for disruption, without a big team, big budget, or a big idea. Which is all they set out to do.

With the heat turned up, ML project owners will run into two formidable obstacles: they'll need talent that isn't available/affordable, and they'll need much more ML training data than they can produce internally

Moving an ML project from pilot to production requires talent beyond the core of data scientists that comprise most early-stage labs. Organizations will realize that they also need machine learning engineers, who can turn ML projects into enterprise software; project managers, with specialized ML skills; and more.Organizations will also quickly realize that the demand for these talents is far outstripping the very limited supply.

At the same time, these teams will discover that they need way more training data than they are able to produce on their own. The POC may have been accomplished with an off-the-shelf algorithm and pre-labeled data. Getting the model to a level of confidence that produces ROI takes much, much  more training data than they have on hand or that they can create with the team and resources they have.

These two obstacles will jeopardize many projects as data science teams try to move them from pilot to production.

While all signs indicate that AI and ML will be key priorities for enterprises in 2019, not all will anticipate the infrastructure and talent needed to successfully turn a POC into a production-ready model. As more organizations reach this stage, demand for AI talent will become even more competitive and they will seek ways to obtain the training data they need to ensure high accuracy in their models.

##

About the Author

Don Roedner is head of marketing for Alegion, a training-data platform for artificial intelligence (AI) and machine-learning initiatives.  In his role, Don is responsible for the company's demand generation, online presence and corporate communications. Don has over 25 years of experience working with B2B software companies throughout his career in capacities ranging from product marketing to corporate communications to demand-generation, for both startups and publicly traded companies.