Computer scientists studying machine learning have trained artificial intelligence using human knowledge of the, uh, carnal variety.
Neural networks are a form of artificial intelligence that “learn” using images or text. They then use that knowledge by mimicking our own brains. They have had a huge year, with major advances made by independent researchers, academics, and huge companies like Google.
We’ve seen neural nets that are trained to identify animals in photos. We’ve seen neural nets trained to describe videos. Last week, Google said it’s using a neural network to generate responses to your emails. Alongside all of this progress, researchers and tinkerers have created dozens of experiments using open-source neural networks, uploading their results to GitXiv.
Today, one of those tinkerers — Samim Winiger, whose work we’ve covered recently — sent along his latest experiment. He used an open-source neural network that was trained on 14 million passages of romance novels by a Ryan Kiros, a University of Toronto PhD student specializing in machine learning. Called the Neural-Storyteller, the network was trained to analyse images and retrieve appropriate captions from its vast store of sexy knowledge, creating “little stories about images,” says Kiros.
And what stories! Winiger fed the network a series of images, and it’s hard to even decide where to begin. How about with this image of Lloyd Blankfein, the CEO of Goldman Sachs?
Not all of the stories (or any of them, really) make perfect sense: What we’re seeing is an artificial neural network struggle to identify objects in a photo, and make links between images and the passages that it’s trained on. For example, take this nonsensical story about Trump:
This image, of sumo wrestlers, gets more right in terms of the shoulder-kissing:
What’s really great about Kiros’ neural network is that he made it open source, and even included directions on how to train it with anything other than romance novels. Winiger, for example, trained it using Taylor Swift lyrics as well.
Sensical? Definitely not. Funny? Yes — that’s actually the whole point. Winiger calls this “computational comedy,” a term he coined to describe using humour to help people understand the extend of emerging AI. As he told me earlier this year, “humour and comedy are a great canvas for education.”
In this case, the idea is to demonstrate both how advanced neural networks have become, in a matter of mere months. But it’s also to show just how far short of our brains they still are. “Neural-storyteller gives us a fascinating glimpse into the future of storytelling,” he concludes. “Even though these technologies are not fully mature yet, the art of storytelling is bound to change.”
If you’re interesting in training your own network — maybe with some Proust this time? — you can check out the code and instructions here.
Lead image: Adam Winsor/Flickr CC.