Enter the fictional world of ganimals

As sea levels rise and habitats are systematically destroyed, species are undergoing extinction at an unprecedented rate. At the same time, new generative AI technologies (such as the generative adversarial network, or GAN) can allow us to imagine new species that have been forced to adapt in order to survive. Hidden within this neural network's imagination, there are millions of these "ganimals" that no one has ever seen before. These ganimals occupy a digital landscape not unlike our own, where attention is short, and engagement is necessary to survive.

To withstand the harsh conditions of the attention economy, ganimals evolve very rapidly. As a result, they occupy a wide range of ecological niches and digital habitats. The data you provide about the ganimals is their "food." Using the feedback you provide, the GAN generates new ganimals by exploring both undiscovered ganimals, and the most highly rated previously discovered ganimals. Thus, by providing feedback on newly discovered species, you can help guide the intelligent evolution of future generations for maximal resilience.

To learn more about the first ganimals ever discovered, click below:

Golden Foofa
Charove
Giuseppe
Baby Oagen


How the platform works

Meet the Ganimals is a collaborative social experiment to discover new species, breed your own, and feed the ones you love. By rating how cute, creepy, realistic and memorable a ganimal is, you can guide its evolution. The cutest, creepiest, most realistic, and most memorable ganimals have a genetic advantage, and breed more often. So over time, the pool of ganimals slowly adapts to the crowd’s opinions.




What is a GAN?

Ganimals are created using a Generative Adversarial Network (or GAN). A GAN is two battling neural networks (the generator, in pink below, and the discriminator, in cyan below) that compete in order to generate new, synthetic instances of data.



The generator takes random numbers and tries to generate new animals from those numbers. The discriminator sees images (both real animals from the training set, and fakes created by the generator). Whenever the discriminator is fooled by the generator, the discriminator updates to be slightly better at discerning fake from real. Whenever the generator makes an image that the discriminator correctly guesses is fake, the generator updates to be slightly better at generating images that look real. So the two neural networks learn together in a beautiful adversarial dance.

What results after training is a generator that can create photorealistic animals. In particular, the generator takes a random number, and converts it to a specified animal. So if you know the number that generates a dog, and you know the number that generates a goldfish, you can use intermediary numbers between the two to generate intermediary imaginary animals. In particular, the number halfway between the two is a perfect blend of the two. This process allows us to “breed” animals to create ganimals.




Beware the training data: The Barracuda Effect

But sometimes there are unexpected consequences of breeding animals, which can be surprising. Although there are no humans in the ganimal breeding pool, we noticed that creepy humanoid characters would emerge whenever we bred an animal with a barracuda.



Since barracudas are ray-finned fish found in tropical subtropical oceans, we were surprised to see this recurring oddity. But when we actually looked at the baracuda in the training data, we discovered something interesting.

Large and strange, Baracudas are prized catches among sport fisherman. So, many pictures on the internet of barracudas include a human holding the fish up, showing off the impressive fish, and giving a comparison of size.



The GAN was trying to generate not only the barracuda, but also the human so often sporting the fish! From this we learned that you can’t blindly trust the training data without digging deeper, and inspecting it yourself. Often times, the data is filled with strange cultural biases that arise from they way it was collected.