Usage

At a general level, creativity is about remixing, reusing, and reinterpreting the world in fresh and unexpected ways. AI can help humans be creative in this way by ingesting massive quantities of examples of something (say, music) and outputting hybrid examples that are both completely new but still similar to the input examples. These kinds of hybrids are termed the latent space of the input data. There are a whole host of ways that an interface could expose this latent space to users, either as a set of dials to tweak, or just by selecting hybrids from the latent space at random. Any time that a user wants to see many different possible versions or options of something, latent space is a powerful tool to provide this capability.

Implementation

One method of generating a latent space is to program an AI to ‘compress’ a piece of data into a smaller piece of data that can still be used to reliably reconstruct the original information. This process is called autoencoding^[1]. Autoencoders take raw unlabeled data, but modern versions can be given a dataset containing multiple ‘instances’ of each example so that the autoencoder can also recognize when two inputs are the same in addition to when they are different. In addition, a variational autoencoder (VAE) attempts to maximize variation between compressed representations so that hybrids can be found more effectively.

Examples

• GrooVAE by Google Magenta: a drum and beat synthesizer that lets users explore the latent space of a dataset of professionally recorded tracks.
• Project Dreamcatcher by Autodesk Research: a computer-aided design (CAD) tool that allows users to specify physical constraints then explore the design space of possible structures that meet those criteria.

Further Resources

• Latent Space Visualization by Julien Despois

Footnotes