Gans

Campers: Iris Xue, Kalkidan Michael, Elizabeth Chu, Priscilla Wu, Cindy Su, and Simon Lee

Faculty: Soheil Feizi

Project Leader: Aya Ismail

GANs are a type of machine learning technique that can be used to produce fake but realistic images. They can often be seen in the real world as deep fakes, where you can make it look like anyone is saying anything. You can also use GANs as a photo editing software, such as Enlight Quickshot, which can accurately guess the background if you delete unwanted things in a photo.

QUESTION:

Can a GAN be used to generate realistic images of fake celebrities?

Answer: Yes! In fact, that is what we focused on in our AI4ALL GANs project, using the Celeba dataset which is full of celebrities.

Tools of the Trade

Action Steps:

We each coded our own DCGANs (Deep Convolutional GANs) and training loops from scratch and trained them on the Celeba dataset.

A GAN consists of a generator and a discriminator. The generator generates fake images and the discriminator distinguishes between real and fake images. The discriminator learns to differentiate real samples from fake samples, while the generator learns how to transform noise into images that look real. Our goal is to train our GAN on image datasets to produce similar images that look real.

For the Celeba dataset we implemented a Deep Convolutional GAN (DCGAN). The discriminator used convolutional layers to produce a binary probability classification, while the generator used transposed convolution layers to upsample random noise into images that had the same dimensions as the Celeba images.

Weight updates alternated between the discriminator and generator. We used the Adam optimizer.

RESULTS:

Some of us were able to get some human-looking final products, but there was also some element of luck involved in training, some having working code but not working photos. It was difficult to find the right hyperparameters that yielded human-looking images. The training process was also long, and took many hours. In addition the algorithms frequently paused or stopped due to usage limits in Google Colab.

Presentation Slides: