Data, the android on the Enterprise in Star Trek: The Next Generation is the first artificial intelligence in our series not written to be a joke. He’s also the most complex, beating out even Bender Bending Rodriguez, although as largely human-like AI they share most of the same processes. What differentiates Data from Bender is his art.
Data’s painting, like Bender’s bad attitude, is something that is easy for viewers to dismiss as unrealistic robot behavior, but before we condemn, let’s think about it. In the late eighties and early nineties, the idea of artificial intelligence generating beautiful pictures was fantastical, but not anymore.
Generative Adversarial Networks (GANs) are an image recognition technique that includes an “adversarial” model whose job is to generate fake images to try to fool the recognition model, thereby forcing it to become better at its own job of recognition.
You read right. In order to really sharpen their model’s skills, researchers make an evil version of the model that tries to thwart it. As long as training lasts, the two models remain locked in a perpetual struggle where neither can come out the victor.
Taking the evil fake image generating model on its own, you have an image generator. They’re all over the Internet now. Some of them even take text as a prompt. Be warned: the quality varies.
That’s… almost a bench. I don’t want to know what happened to the woman.
This one gives me a new mutated bear each time I hit “submit.”
If that makes you laugh, try these non-interactive examples reportedly generated by Google’s Imagen. Yes, those captions are the prompt and the AI generated the images above them to match.
It’s hard to dispute that computers can now create strikingly realistic images. Whether doing so makes them “creative,” well, that might come down to Picard’s proof by unknowability argument in Season 2, episode 9. To paraphrase his defense of Data’s sentience, the images generated are unique and evocative. Let anyone in the comments show me a means of measuring the soul of a work of art.
Name: Lieutenant Commander Data
Origin: Star Trek: The Next Generation
Likely Architecture: Reinforcement Learning, Convolutional Neural Networks for vision processing, Generative Adversarial Networks to make art, and Transformers for Speech and Language. Neurosymbolics for simple reasoning.
Possible Training Domains: Life in Noonian Soong’s lab, life on the Enterprise.
I take requests. If you have a fictional AI and wonder how it could work, or any other topic you’d like to see me cover, mention it in the comments or on my Facebook page.