Loops in BayeuxSpace

A collaboration with Tom Chambers of Random Quark

These videos are generated by a neural network trained on images of the Bayeux Tapestry.The process of changing a trained network’s input parameters and observing the output is often described as exploring a space. In that sense these animations are looping paths within that space.

Extending the Bayeux Tapestry is a task that has been attempted before by human minds far more informed than any of the machine learning models we experimented with. The tapestry has been repeatedly revived and claimed and drafted in to bolster various causes at different points in European history (and so it continues), but the human tendency to identify with one side or the other isn’t always helpful.

Algorithms can generate new imagery naively, on a purely visual level, indifferent to the features we’d see as being loaded with meaning or constituting narrative, and unthinking when it comes to context. We thought it was possible that the machine’s distance from the subject matter could give it some apparent insight – generating images of meaningless conflict without heroes or villains, justice or honour.

The entire tapestry constitutes a very small dataset in machine learning terms. Our trained network suffers from “overfitting” – rather than demonstrate a general understanding it tends to generate images that can be traced back to particular scenes in the original. However, it still produces some interesting intermediate states in the space between those scenes.