A group of software engineers from Google's AI department shared their research on the Artificial Neural Networks and its image recognition capabilities.
When the team requests its Artificial Neural Network to identify a pattern on an existing picture after encoding it with several images, the system tends to interpret it with abstraction. The engineers call the technique Inceptionism.
The process is similar to how humans see different shapes in the sky while cloud watching. With the range of our imagination, we often perceive actual objects in the clouds such as cars, faces or flowers. Google is developing a very advanced type of vision system that will allow an Artificial Neural Network to identify the picture it is looking at and distinguish from other objects, such as an ant versus a starfish.
Google explains that the structure of its neural network is composed of layers of artificial neurons that can be as many as 30 at once. When an engineer encodes a picture through the network, the first layer of neurons will detect any low-level information such as the edge of the picture. The next layer of neurons will identify the shapes on the picture, until the final layers of neurons are combining all the information gathered by the other layers to complete its interpretation of the picture.
Google "trains" these Artificial Neural Networks by filling it up with millions of pictures, and the engineers can even tweak the system to focus on a specific type of photos such as animals or trees. The engineers recently discovered how the system has advanced when they requested it to generate images of certain objects starting from just noise or a bunch of dots with colors.
As the tests continue, the engineers were amazed with the outcomes and asked the system to enhance the functions of the layers of neurons. For instance, the layer in charge of identifying the edges of pictures were asked to generate objects on the pictures, and the processed images now look like the same picture with different colors and shapes.
In one case, the engineers ran a picture of clouds to a neural network trained for animals and the results of the Inceptionism were amazing. The system has combined different animals within animals within the clouds.
Interestingly, the engineers state on the blog that they also placed the network in an endless loop for it to generate an image on an already processed image and the network produces a mixture of abstract images that can be considered works of art.
Check out the whole gallery of Inceptionism by Google.
