Google Has Made It Simple For Anyone To Tap Into Its Image Recognition AI (gizmodo.com)

An anonymous reader quotes a report from Gizmodo: Google released a new AI tool on Wednesday designed to let anyone train its machine learning systems on a photo dataset of their choosing. The software is called Cloud AutoML Vision. In an accompanying blog post, the chief scientist of Google's Cloud AI division explains how the software can help users without machine learning backgrounds harness artificial intelligence. All hype aside, training the AI does appear to be surprisingly simple. First, you'll need a ton of tagged images. The minimum is 20, but the software supports up to 10,000. Using a meteorologist as an example for their promotional video was an apt choice by Google -- not many people have thousands of tagged HD images bundled together and ready to upload. A lot of image recognition is about identifying patterns. Once Google's AI thinks it has a good understanding of what links together the images you've uploaded, it can be used to look for that pattern in new uploads, spitting out a number for how well it thinks the new images match it. So our meteorologist would eventually be able to upload images as the weather changes, identifying clouds while continuing to train and improve the software.

Comments

Re:They want the crowd to train their machine

[TheRaven64]

I wonder if that's actually true. On the one hand, they get a lot of training data for free. On the other hand, how well are the protected against adversarial training? If a botnet signs up for a million accounts, submits 20 photos with correct descriptions for 19 and incorrect ones for the remaining one, is it possible for the attacker to teach the Google algorithm some nonsense things? What if I'm more subtle and upload 100 pictures of dogs and tag them as dogs, but the pictures are carefully crafted to share more characteristics with other pictures of cars than of dogs? Can I train it to recognise all cars as dogs (last year researchers were able to alter a picture of a dog or a car by one pixel and have this system recognise it as the other). A system like this is only as good as its training set - remember that this is the same system that decided that all black people were gorillas because the only black face in its original training set of photos from Google employees was a gorilla at a zoo.