Artificial Neural Networks use well known mathematical methods to help computers learn to recognise and interpret patterns from experience, allowing them to classify images and recognise speech. Layers of decision making ask what structure something has and then assesses it in more detail finally deciding what things the computer has been shown before this is most like.
Running the process in reverse produces some surprising results and shows just how the computer sees objects in the world. It can also highlight when it has been misled by the images fed for a particular object; that the computer’s “training” has gone awry.
Read more on the Google Research blog.