Google has recruited its in-house machine learning framework, TensorFlow, to help train additional spam filters for Gmail users. With the new filters in place as of last month, the company claims Gmail is now blocking an extra 100 million spam messages every day.
In the context of Gmail’s 1 billion-plus users, this isn’t necessarily a huge gain — it works out as one extra blocked spam email per 10 users — but Google says Gmail already blocks 99.9 percent of spam, so working out what constitutes that last sliver of a percentage is hard.
“Getting the last bit of incremental spam is increasingly hard.”
“At the scale we’re operating at, an additional 100 million is not easy to come by,” Neil Kumaran, product manager of Counter Abuse Technology at Google, tells The Verge. “Getting the last bit of incremental spam is increasingly hard, [but] TensorFlow has been great for closing that gap.”
Gmail has been using AI in addition to rule-based filters for years. While rule-based filters can block the most obvious spam, machine learning looks for new patterns that might suggest an email is not to be trusted. Algorithms trained in this way balance a huge number of metrics, everything from the formatting of an email to the time of day it’s sent. TensorFlow, says Kumaran, makes managing this data at scale easier, while the open-source nature of framework means new research from the community can be quickly integrated.
TensorFlow was launched by Google in 2015, and it has become an incredibly important part of its AI business. It’s a free machine learning framework that allows developers to create AI tools for a huge range of tasks. Fans praise its flexibility and capacity to scale, and, of course, it works seamlessly with Google’s other AI services, encouraging users to buy computing power from Google as well as off-the-shelf vision and speech algorithms.
Google says integrating TensorFlow into Gmail will also allow it to better personalize spam filters. This process has been taking place for years, says Kumaran, with Gmail looking for certain signals from users about what they judge to be spam, but TensorFlow is “turning those signals into better results.”
“There’s no one definition of spam out there,” he adds. But AI could help work out the best definition for you.