Pre-trained machine learning models are good enough for many use cases, but to get the most out of this technology, you need custom models. Given that it’s not exactly easy to get started with machine learning, Google (and others) have opted for a hybrid approach that allows users to upload their own data to customize the existing models. Google’s version of this is AutoML, which until now only provided this capability for machine vision tasks under the AutoML Vision moniker.
Starting today, the company is adding two new capabilities to AutoML: AutoML Natural Language for predicting text categories and AutoML Translation, which allows users to upload their own language pairs to achieve better translations for texts in highly specialized fields, for example. In addition, Google is launching AutoML Vision out of preview and into its private beta.
Rajen Sheth, the director of product management for Google Cloud AI, said that this extension of AutoML is yet another step toward the company’s vision of democratizing AI. “What we are trying to do with Cloud AI is to make it possible for everyone in the world to use AI and build models for their purposes,” he said. For most of its customers, though, pre-trained models aren’t good enough, yet for most businesses, it’s hard to find the machine learning experts that would allow them to build their own custom models. Given this demand, it’s maybe no surprise that about 18,000 users have signed up for the preview of AutoML Vision so far.
“Natural language is something that is really the next frontier of this,” Sheth noted when he discussed the new Natural Language API. “It’s something that’s very useful to the customers. Because more than 90 percent of our customers’ information within their enterprise is unstructured and free information. And a lot of this is textual documents or emails or whatever it may be. Many customers are trying to find ways to get meaning and information out of those documents.”
As for AutoML Translation, the benefits of this kind of customization are pretty obvious, given that translating highly specialized texts remains the domain of experts. As an example, Sheth noted that “driver” in a technical document could be about a device driver for Windows 10, for example, while in another text it could simply be about somebody who is driving a car (until computers take over that task, too).