Google introduces Cloud AutoML for employing machine learning without experts
The first offering in this new service, AutoML Vision, can be trained via a graphical interface with a few dozen images.
Machine learning has become essential to digital marketing because it allows the generation of predictive models from past data. These predictions allow marketers to target the right offer at the right time, and they help systems recognize things in the real world.
Today, Google Cloud is taking another step toward making machine learning accessible to every level of business.
It is introducing its Cloud AutoML service, which provides access to custom machine learning models that the company says can be trained “with minimum effort and machine learning expertise.”
This complements the Google Cloud Machine Learning Engine, released last year, that lets developers with machine learning expertise create models for any kind of data. As part of this service, Google has already made available via APIs pretrained models in Vision, Speech, Natural Language Processing, Translation and Dialogflow (for voice- and text-based dialogues) for business applications.
The idea with Cloud AutoML is that developers who don’t have a lot of experience with machine learning can still join the game.
It assumes that the business has access to developers, of course, but companies can take advantage of the opportunity without hiring machine learning specialists. A drag-and-drop user interface lets developers quickly start the model training with sample data.
The first product from this new service is AutoML Vision, designed for training self-learning machine vision models to meet a brand’s own needs, employing Google’s home-grown image recognition tech.
Google says initial tests show that AutoML Vision is more accurate at image recognition than generic machine learning models. A simple model for use in an application can be created in minutes, the company says, and a full model only takes a day to complete.
A clothing company, for instance, might teach Vision to recognize different kinds of sweater types by their neckline. Google says that, in some cases, the brand can teach Vision with as few as several dozen photographic samples, after which Vision gets the idea and will teach itself to recognize sweater types.
Google points to the use of AutoML Vision by Urban Outfitters to automate the recognition of product characteristics for recommendations to consumers, by Disney to recognize its characters through visual search and by the Zoological Society of London to automatically recognize animals photographed in the wild through camera traps.
AutoML itself is the result of Google’s research, where a neural net continually generates a “child” machine learning model with a specific architecture for a specific task, such as visual recognition. Feedback on accuracy and other factors informs the next child model and the next, and so on for thousands of generations of new model architectures.
Google says it has applied this kind of evolutionary improvement of machine learning models in two areas: image recognition and language. In both cases, the company says, the best automatically generated models have an accuracy comparable to the best models designed by machine learning experts. In fact, in some cases, the machine-generated versions appear to have created useful new architectures.