Using Machine Learning to write my museum conference keynote title

I was invited to give a keynote talk at the wonderful Australian Museums and Galleries Association annual conference, in Alice Spring, in May 2019. Because the talk was, in many ways, about the false utopia of shiny new technologies in the museum world, I felt it was appropriate to unnecessarily (and self-indulgently) utilise an en vogue technology while developing the talk. I feel some confidence in exploring this idea space as I’ve produced more than my fair share of high-gloss tech projects for museums over the last 10 years.

I chose the much hyped technology Artificial Intelligence, and more specifically a text-based Recurrent Neural Network (RNN), to create the title for the talk and for each of the main sections within. I trained an RNN on the available history of presentation titles for this conference, going back to 2001 and including this year (thanks AMaGA!). There is over 700 titles, which is not enough to make an intelligent machine, and I deliberately didn’t train it on Australian English words or grammar. The AI only speaks the language of Australian Museums and Galleries Association Conference Session Titles.

The process:

1. Scraped the data and extracted the titles
2. Downloaded textgenrnn from Max Woolf, a data scientist at Buzzfeed, which meant updating phython and pip and installing Tensorflow (from Google and Udacity) which does the underlying grunt work
3. Trained textgenrnn on the data and repeated that training until I was getting good results — i.e. about 5 times
4. Selected the right temperature, which just means how much freedom to give the RNN
5. Took all the output and piped it into a raw text file
6. Wrote a script that threw away results that had lots of ‘and’s or were too long, and selected candidates that had words that are part of in my talk such as technology, interne, network, change, collections, digital, design, interface, experience, people, community, future and similar
7. Picked the best of those that remained

It took about 3 episodes of Russian Doll to complete, including the time the RNN took to run.

screenshot of textgenrnn running on the command line

The output:

In the end, I selected my favourite title option for AMaGA:

The Museum Program and the Survival of the Future

A few others are used to introduce sections:

• The section on museum websites became “Learning net protocolity”
• Museum apps became “Using Museums: A Connection or Call”
• Interactive exhibits became “Australian technologies where sustaining lights in names for the halon public exhibit”
• The AR/VR/XR section became “Institutions in the Post Museum”
• And the section about the use of AI in museums got, perhaps, the silliest “Something the partnerships of changing museums and how are we need to make them?”

However most of the ~500 titles options the RNN generated were discarded. Fully 50% of them were gibberish.

Here are a few other candidates, read through them and ask yourself if you’d go along to a session with one of these titles?

• Beyond Volunteers: struggle and change for the centre interface
• Design and the People of The Australian Museum of Australian Museums
• The City of Myth Climate Stories
• The Transition To Museum From Collecting Connections
• Memories from the museum through a collection
• Hillas: The National Forgotten Australian Museum of research
• The National Museum of Australia & the Australian Museum of Victoria: the future of collections in the 21st century
• A collections of collections in exhibition: a human audience
• Community Exhibitions of Community
• Public Curators: A story of the museum botanic community project and collections and research and learning and communities and in the 21st contemporary community museums in a project and community to provoke to determine to a program at the Australian Museum of Australia

In closing:

I believe that AI’s aren’t, in themselves, creative. Something I’ve argued in my capacity as Professor of Practice at UNSW Art & Design at events such as Sydney Design Week and the UNSW Creativity and AI Symposium. We need new language to describe the novel output of these machines, language that isn’t so closely tied to human intelligence or creativity.

AI is often and effectively used as meta-commentary on itself and the field it’s being trained on, as it was in this case. Where AI is most exciting for museums, I believe, as a form of human thought augmentation — an automated brain poking machine that can throw ideas at its human partner for them to consider.

The recent experiments at The Barns, The Met and MoMA offer some signs of productive use of AI/ML in museums, however, I believe there is much to learn and much to do before we see these technologies providing a genuine value add for museums.

—

Props to Greg, Max, Beth, Dan, Janelle, Shane, and all the others who’ve done this sort of experimentation before me.