Google knows what the dreams of a neural network are

Artificial neural networks Google created to simulate the human brain. This technology allows you to recognize and analyze different images. One day the developers had a curious question: what would happen if the robot could dream? Such a strange question did not arise from scratch. It is part of the project to create images Deep Dream.

"Deep Dream"

The developers put a specific goal before the software. However, this goal was not the re-creation of dreams. Experts asked the neural network to change the image based on the original photo by overlaying several other layers on it. As it turned out, the software is easy to learn. Thus, the program was able to improve the detection functions of the specified models.

Training

To improve the functions of artificial neural networks, developers have passed more than one million images through the computer. It was a painstaking and time-consuming work, after all, after each proposed picture, the engineers forced the machine to emphasize the object recognized in the image. The neural network itself consists of several layers, and a more accurate interpretation of the search depends on their level and status. For example, an output layer is responsible for recognizing individual objects.

Hallucinogenic picture quality

After increasing the recognition of specific objects on the image of the neural network, more complex work was to be done. Engineers gave the machine the task to create images of certain objects, among which were a dog, fork, starfish, banana and other objects. This step has fully justified itself. And let the robot dreams have a hallucinogenic quality, the human eye can recognize the given images.

The ultimate goal of the project

Google wants to improve neural networks to a state where it would be possible to recognize non-existent details in a common picture. We can say that the engineers managed to look into the subconscious of artificial intelligence. This happened when developers started uploading images to the top layer of the neural network, one that learned to recognize individual objects. So, for example, the given parameter "the outline of the dog in the clouds" made the network model from the clouds of the dog. And with each subsequent download the result went out better and better.

Thus, "Deep Dream" gave the computer the ability to independently change the image parameters. And this allowed us to recognize objects that are not contained in the image. And now, at the request of "cloudy sky" the network produces surprisingly strange dogs and snails.

Conclusion

Methods used by scientists during the implementation of the project, help to understand and visualize how neural networks are able to perform complex tasks for classifying objects. This led to the improvement of the network architecture and allowed to control the steps of the learning process.