Google DeepMind: an advance in images with AI

Google’s AI has once again shown its advance on images, DeepMind is now able to create images that never existed. For example, DeepMind can generate dog pictures, mushrooms or spaghetti dishes that do not exist as shown in the pictures below:

It is therefore very difficult to distinguish these images from real images.

To accomplish this feat, DeepMind used an unsupervised learning algorithm, Generative Adversarial Networks (GAN). In this system, two neural networks work in parallel, the first network creates the images and the second network confronts them with the real images. If the image is validated, the algorithm continues its image design.

For this algorithm to create such real images, Google has used its gigantic computing power, Google teams are also talking about BigGAN.

Indeed, the BigGAN will use 158 millions of parameters for the generation of images and 2048 samples for each comparison. The images thus generated obtained an IS (Inception Score, which is a score that measures the diversity of the generated images) three times larger than the previous experiments and a FID (Frechet Inception Distance) which estimates the distance between the false images and the real images) twice as small.

The goal of this research by Google DeepMind is to be able to generate images from simple words. Examples of applications could be an AI capable of producing a film from a script, a book or automatic text illustrations.

However, generating images that do not exist tends to think that a generalization of false information with fake photos could occur, but for now, there is no such risk because the generation of such images requires gigantic computing powers.

For example, it takes 24 to 48 hours of calculations with a module of 512 TPUv3 Google (Tensor Processing Unit) to generate a single image of 512 pixels wide. As each TPU consumes 200 Wh, each image is therefore between 2,457,600 kWh and 4,915,200 kWh. This represents the average electricity consumption of a French household for six months or the electricity consumption of the city of Cleveland (386,000 inhabitants) for an afternoon.

To conclude, it is thanks to a huge computing power and networks of generative antagonist neurons (GAN), that it is possible to create photos from scratch. This area of ​​research raises fears about the possible manipulation of false information on a large scale. However, this technology is not accessible to everyone.