Artificial intelligence already affects most of our lives in a variety of ways, like helping you drive, sprucing up our food Instagram pictures, and even helping to diagnose diseases. AI crunches data differently than humans, which is why communicating with it in our weird, inconsistent language can turn up such wacky results. The latest collection of strange AI creations comes from an Attentional Generative Adversarial Network that translates typed words into a visual image.
Cristobal Valenzuela, who builds machine learning tools, created the site to demonstrate how AI can parse words and try to convey their meaning in a visual way.
It was first spotted by the excellent AI Weirdness blog, which succinctly describes the entire process as like “a visual chatbot in reverse.” Instead of trying to tell you what’s in a picture, the algorithm tries to generate a picture from what you tell it. The concept is based on research from a paper published last year called AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks
The bot draws from a collection of objects on which it was trained. As a result, it’s much more accurate with everyday objects that easily translate into a picture. Things get a lot more abstract when you start feeding it nebulous terms, complex concepts, or, my personal favorite, wacky nonsense.
We fed some science terms into the tool to see what it could do. The results are predictably bizarre, at least to my non-robot brain.
You can try your own phrases here, but be warned that heavy traffic from Twitter has made the service a little unreliable.