During my time at BBC R&D, Cassian Harrison, Controller of BBC Four, challenged us to produce a one-hour documentary programme, assembled from existing archive clips, entirely using machine learning techniques:

This was an interesting exercise in testing the limits of computational creativity, and in the use of machine learning on media to educate a non-specialist public about the technology, it’s strengths, weaknesses, and possible risks.

Alongside my editorial colleague Caroline Alton, I led the team which put the film together: investigating, devising and implementing the machine learning technology which produced the film itself.

We used a variety of different approaches to do this, including computer vision and natural language processing, techniques, as well as adapting recent deep neural network-based sequence learning techniques to model the sequence of transmissions in our film. All the details are availble in my post on the BBC R&D blog!

The programme was broadcast on BBC Four in September 2018. We also published a paper on the project (in collaboration with Daniel Chavez Herras from Kings College London) at the Digital Humanities Conference 2019.