Machine learning gets a boost

New automation platforms like AutoML promise to make machine learning processes more scalable and cost-efficient

automated machine learning
  • As more companies adopt AI tools, they are looking to make machine learning processes more efficient and cost‑effective
  • New automation platforms such as AutoML free data scientists from having to train ML algorithms manually
  • Companies still need human specialists to ensure accuracy of data and results

Many companies have invested in artificial intelligence solutions. Only 51% of them hire their own specialists to train and test algorithms, according to a 2018 survey by O’Reilly Media. The rest either outsource machine learning needs to outside vendors or license AI algorithms off the shelf.

Automated machine learning, or AutoML, makes it possible to overcome many process automation hurdles of machine learning without human attention. It also frees up data scientists to concentrate on other priorities, and makes machine learning accessible to organizations that lack dedicated AI teams.

Consider one common machine learning task: classifying a large set of images. In traditional machine learning, a data scientist must select the best model to categorize the images, define its limits, check outputs, and repeat the steps until the model achieves the desired results. AutoML automates that time‑consuming process.

“With AutoML, you provide your input data and what you want it to predict, and it will iterate through all the available machine learning models to choose the best ones,” says Kai Wombacher, a data scientist at Zylotech, an analytics company that uses automated machine learning models for customer analytics.

Time savings and greater access to ML

For organizations with full‑time data scientists, freeing them from repetitive, laborious tasks translates into cost savings. “Instead of spending time selecting and tuning models, there’s more time for data scientists to ask questions and tackle new issues,” Wombacher says. “Freeing up their bandwidth means they can move on to developing new models.”

The other draw of AutoML is its accessibility to organizations without robust AI teams. Google, which developed and launched its own AutoML cloud solution, says the technology can “enable developers with limited machine learning expertise to train high‑quality models specific to their business needs.” Data scientists are among the highest‑paid personnel in IT. With the help of AutoML, many companies can develop strategically useful AI capabilities for less.

Companies looking to add automation to their machine learning processes have limited options at the moment. They can try off‑the‑shelf solutions from companies like Google, experiment with open‑source options, or build their own.

While fewer than 5% of companies in the O’Reilly survey currently use AutoML tools, a few early adopters have seen success. The Zoological Society of London recently used Google’s Cloud AutoML platform to automate the process of analyzing and annotating millions of images captured by camera traps in the wild, a job previously handled manually.

Zylotech uses a similar, open source tool to automate training of its customer analytics models. Those models now predict customer churn with 95% accuracy. The biggest benefit is that the tool frees up Zylotech’s data scientists to develop the logic for additional models.

Early days for AutoML

AutoML doesn’t make data scientists obsolete. For it to work properly, companies still need people who are familiar with managing and manipulating data to ensure that data inputs are accurate and not corrupted with extraneous information, which can throw machine learning models off course. “Maybe you need fewer data scientists or less bandwidth from a data scientist, but you still need them,” Wombacher says.

It’s also worth noting that AutoML doesn’t necessarily speed up the process of training machine learning models. Computer run times can be just as long with AutoML as with manual testing.

“While it saves you on the development of the model, the output demands a lot of time,” Wombacher says. Companies also still need to compile datasets, understand how machine learning works, and test hypotheses against outputs.

Another challenge: AutoML is emerging tech that has not yet been standardized. “Every company ends up working on their own version of AutoML,” says Andrey Turovsky, head of AI initiatives at ZoomInfo, a contact‑database company. “I don’t think it’s far enough along yet that we’ll see one general‑purpose platform for everyone without having to customize and get optimal results.”

ZoomInfo is in the early stages of developing its own automated platform for machine learning. “It’s a lower level than a general‑purpose AutoML platform,” Turovsky says. “But our goal is to automate a lot of the steps that are required in creating a prediction model, and be able to use it in production.”

While use of AutoML today remains limited, practitioners agree that it will eventually become as fundamental to machine learning as the algorithms themselves.