Google Brain: A New Breakthrough in Artificial Intelligence by Google. SAN FRANCISCO. Explore the cutting-edge technology behind Google Brain. On February 8, in a significant development, Google announced a major breakthrough in its artificial intelligence department. Google Brain has unveiled a new software that can convert mosaic-like source images into high-definition, code-free images using Pixel Recursive Super Resolution technology.
Take a look at the image below. The left column displays an 8×8 resolution encoded image, the middle column showcases the image restored by Google Brain’s software, and the right column shows the original image. The result is remarkably close to the original picture.
Typically, you cannot generate a more detailed image than the source picture. How does Google Brain achieve this? It intelligently combines two neural networks.
Working Principle of Google Brain
The first part is the conditioning network, which compares an 8×8 image to a high-definition image, reducing other HD images to 8×8 resolution and then matching them.
The second part is the prior network, which utilizes PixelCNN (pixel-based neural network) to add genuine HD details to an 8×8 source image. The prior network has been trained on various images, such as those of celebrities and bedrooms. When the source image needs to be enhanced, it identifies matching pixels from its database and adds them.
For example, if the source image contains a brown pixel at the top, the prior network might interpret it as an eyebrow, which leads to image enhancement, adding brown pixels in an eyebrow shape.
To create the final code-free high-definition images, the outputs of these two neural networks are integrated. The end result may sometimes contain spurious new details.
Google Brain’s high-definition enhancement technology has achieved significant success in real-world testing. When the team presented authentic HD celebrity photos and computer-enhanced images to human observers, 10 percent of the observers were deceived (50 percent achieved a perfect score). In the case of bedroom photos, human observers were misled 28% of the time. These scores far surpass those of conventional interpolation techniques, which do not deceive any human observers.
Limitations of Google Brain’s Image Restoration
It’s important to note that the images restored by these algorithms are not true representations but are based on “educated guesses.” This technology raises intriguing questions, particularly in the realms of surveillance and forensics. It can enhance blurry suspect photos, providing some assistance to law enforcement.
Google Brain and DeepMind are two leading institutions in deep learning under Alphabet’s umbrella. Google Brain has recently published some fascinating research, including the development of proprietary artificial intelligence encryption algorithms. DeepMind gained fame for developing AlphaGo, which defeated the world’s top Go players last year.