“Our goal is to create applications that can see, hear and understand,” says Jeff Dean, senior fellow at Google and one of leaders of the firm’s machine learning strategy.
Artificial intelligence is a clear priority for Google. Last week, its AlphaGo program defeated world Go champion Lee Sedol, while its DeepMind arm is considered to be at the forefront of deep learning research. Machine learning also underpins many of its well-known products, from YouTube to Google Translate.
And now Google is increasingly opening its technology up to other businesses as it attempts to grow its cloud operations. (See also: what is a graph database?)
At its Cloud Next conference Google unveiled the Cloud ML service which helps machine learning engineers build “sophisticated, large” models based on its TensorFlow deep learning library, which was opened sourced last year.
Google also added to its list of ‘pre-trained’ machine learning platforms aimed at programmers without data science skills. Cloud Speech allows developers to convert audio to text by “applying powerful neural network models in an easy to use API”, the company says, and joins its Translate and Cloud Vision APIs.
How can businesses use machine learning?
Speaking to press at Google’s cloud user conference, Dean said that its machine learning tools will “allow other companies to build the same kind of understanding and insight from their data” that Google has achieved internally.
This could be applied in retail, for instance.
“For example, if you have some data about customers and transactions and you want to predict something about it, that is a very general setting that can be used in all kinds of different enterprises,” Dean said.
“So you could predict which customers are going to buy more than a hundred dollars of stuff and send them an offer.”
Online retail logistics firm Ocado Technologies already uses machine learning algorithms in various ways across its business, such as powering its robotic systems. General manager James Donkin, says that the company has been evaluating TensorFlow to help it build new algorithms, and would welcome the ability to move workloads to the cloud.
“We really like the fact that we can run it locally, but if we want to use it on a larger data volume it is easy to move it to the Google cloud to run the same TensorFlow code,” he said, adding that Google’s decision to open source the framework is “a big plus”. “We don’t like black box technologies that we can’t look inside," he said.
Democratising machine learning access
Dean says while the TensorFlow and Cloud ML tools are aimed at more sophisticated machine learning use cases, the cloud-based APIs cater to a wider audience of developers.
One example is online hosting firm, Wix, which is using the Cloud Vision API. The technology is based on the image recognition that allows Google Photos to categorise photos of mountains or beaches, for example.
Dean explained: “Wix allows people to easily use websites and clients upload imagery. They want to be able to understand what kind of imagery it is so that they can suggest the most appropriate content of that website.”
He added that there are also some that “can make pretty good use of the already trained APIs that we have released”.
Ocado’s Donkin says that the ease of use of the pre-trained APIs means the company could extend access to machine learning tools from its team of data scientists to developers.
“One of our goals is to move beyond just data science using machine learning,” he told ComputerworldUK. “So some of the new Google APIs - where normal developers can use machine learning - [could be] something we would do.”
“We are looking at the structure we need to do that. So you have the PhD specialist groups and you have software engineers, and how we can organisationally let that knowledge spread out will be.”
Like many of Google’s cloud services - such as BigQuery and Hadoop-based DataProc - its machine learning services originate from technologies that it builds to run its massive operations. This includes Maps, Gmail, Android, as well as its robotics research.
“Machine learning is now used in lots and lots of Google products,” said Dean. “Some under the covers - so YouTube has lots of machine learning in it but it is all in the recommendation engines.”
Google plans to continue to offer up its AI tools to its enterprise customers.
“We think there is a lot of opportunity to introduce more and more of these pre-trained APIs, and example models for particular kinds of settings. So recommendations might be a particular type of model,” Dean said.
That said, there are some areas which Google is unlikely to make available to a wider audience, in particular the algorithms helping power its search business.
“It is generally a business decision about which ones make sense to release,” Dean explains. “We have been pretty open about the ones we have released. I would say things like search ranking we will probably not release but other than that most things [we would consider releasing].
“We want people to use this to build their own cool products as well.”