AI SCIENTIST IAN GOODFELLOW QUIETLY posted these five images to Twitter shortly into the new year. His tweet had no other statement but to append links corresponding to each picture.
Keep in mind: none of these ‘people’ exist, except as rendered by a machine. It uses a new form of AI created by Goodfellow and his colleagues known as GANS and first shown in 2014. GANS stands for Generative Adversarial Networks.
GANS turns random data into images by putting two neural networks in competition with each other. The original paper by Goodfellow and his fellow researchers compared the process to a counterfeiter. One image is the real thing and the other a representation, constantly being refined to match the real thing, working until the police can’t tell whether it is seeing fake or actual currency.
With these two neural networks at work, one generates a result (the generator) while the other discerns how closely that data-generated result matches a known result (the discriminator, aka the classifier). This back-and-forth continues (the adversarial part) until the generated ‘made up’ image is deemed to be indistinguishable from the real thing by the discriminator. It is as if to say, ‘yes, that image is a human face because it characteristically resembles all the other faces I know.’
The difference from counterfeiting is what happens next. There is only one kind of $100 note, but an endless variety of faces. Once the GANS system can produce a credible image of a face, it can introduce any number of variations of faces that are equally credible. Variables are things like eye colour and shape, hair colour and length, as well as bigger differences in sex, age, skin colouration, and more.
OUR TAKE
- These five images are a 5,000 word equivalent of how far GANS technology has come in four years.
- GANS presents a path for computers to be creative, generating outputs heretofore unrealized, and possibly unimagined, by humans.
- On the flip side, its life-like imagery can produce means of deceit that is harder and harder to detect.
Here are the supporting papers for each stage of development from Goodfellow’s group:
- Generative Adversarial Networks
- Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
- Coupled Generative Adversarial Networks
- Progressive Growing of GANs for Improved Quality, Stability, and Variation
- A Style-Based Generator Architecture for Generative Adversarial Networks
SEE ALSO
- Talks about AI | GANS: How machines convert data to images
- AI in the newsroom | Novel AI system can draw caricatures | THE INDIAN EXPRESS