Tech

Check out these fake album covers created by artificial intelligence

by Jeremy Mersereau

And so it begins: first machines are learning how to generate fake album covers, next they’ll be enslaving us all to their unknowable whims… and that’s only if we survive the great Flesh Purges of the coming decades. For now though, we can still complacently ooh and ah when an AI does something cool like learns how to generate convincing fake bedrooms or album covers, the equivalent of a wolf puppy learning to sit before it grows up and mauls you. Let’s enjoy it while we can!

Machine learning developers and researchers Alec Radford, Luke Metz, and Soumith Chintala wanted to push forward the learning aptitude of CNNs, or convolutional neural networks, especially in regards to unsupervised learning. To achieve this, they created and coded a new class of CNNs, DCGANs (Deep Convolutional Generative Adversarial Networks), which contain ‘architectural constraints’ that aid them in effectively determining a ‘hierarchy of representation’ when given image datasets to learn from.

I’m pretty sure this is how Skynet started. Note to coders out there going forward: let’s try not to include the words “adversarial networks” in our new creations, OK?

The trio’s paper is under review for the 2016 International Conference on Learning Representation, but they’ve already provided example images and the source code over at GitHub. Here’s what the DCGANs came back with after five passes through a dataset of bedrooms. These images are completely generated by the network:

Pretty creepy, particularly if you have a phobia of geometrically-improbable beds. But now the main event: what does the DCGAN come back with when provided with image sets of classic album covers? Click/tap the image below to enlarge:

Welp, that’s it. Close it down, we’re done here, I fully expect our new AI overlords to do better than we ever could, anyway. I could be a great Renfield for you in the meantime, my DCGAN master!