Google has updated its Google Photos app on Android with a new option that lets users tell the search giant about the contents of their pictures. By labeling these images, Google can improve its object recognition algorithms, which in turn make Photos more useful. It’s a virtuous cycle of AI development best deployed by tech giants like Google which have lots of data and lots of users.
This isn’t an unusual practice at all. Machine learning systems don’t just learn by themselves, and the vast majority of these applications need to be taught using data labeled by humans. It’s the same reason that CAPTCHAs ask you to identify cars and motorbikes in images. By identifying these objects you’re training AI to do the same.
The feature appears in the most recent version of Google Photos. Just tap on the search button in the app’s menu, scroll down, and you’ll see an option to “Help improve Google Photos.” As reported by 9to5Google, click on it and you’ll be presented with four tasks: to describe your printing preferences for photos; your preferred collages or animations; to identify which photos belong to which holiday events (eg Christmas or Halloween); and to identify the contents of photos (“Name the most important things in this photo”).
As Google explains on a help page about the feature: “It may take time to see the impact your contributions have on your account, but your input will help improve existing features and build new ones; for example, improved suggestions on which photos to print or higher quality creations that you would like. You can delete your answers at any time.” (To do so, tap the three-dot menu at the top right of the screen and hit “Delete my answers.”) At the time of writing, it seems the update is available only on Android, not iOS.
Although this looks to be a new addition to the Google Photos app, the underlying software is much older. The process is powered by “Crowdsource by Google,” a crowdsourcing platform that the company launched in 2016. It gamifies data-labeling, letting users earn points and badges by completing tasks like verifying landmarks, identifying the sentiment of text snippets (is a review positive or negative, for example), transcribing handwritten notes, and other similar jobs. To be clear, though: users don’t get any real rewards for their work beyond virtual kudos from Google.
It’s worth remembering all this when using Google’s whizzy machine learning products: they wouldn’t be half as good without humans helping teach them.