We Got an Artificial Neural Network to Name Random Objects in the VICE Office and It Did OK
Finally someone is putting puppyslugs to some good use.
A neural network looking at an image made by a neural network *mind explodes.* Screencap from AI•Scry app taken by the author
Normally when thinking of artificial neural networks (ANNs), the mind's eye explodes with predictably trippy images of puppyslugs and other computer-generated mutants originating from Google's so-called Inceptionism program. But a new app from Disk Cactus called AI•Scry is a great example of the other applications for ANNs in our every day life. It uses your camera to look at normal objects, then uses a database of images to guess what it's looking at. This is similar to the tech Google's DeepMind program AlphaGO used to beat champion Go player Lee Se-dol at the ancient Chinese board game, but rather than plotting out Go strategies, it guesses whether that's a cup of coffee sitting on your desk, or a cat sitting in a bowl.
Because Disk Cactus is a zany company—you might remember their very necessary invention, the Emoji Keyboard—AI•Scry includes an "Attention Aperture" slider which messes with the AI's guesses. The app describes itself as a "remote viewing application powered by an alien psyche," and that alienness shows with the slider, which can take a tree from "tree with green leaves" to "black kits and grime and grime attached in a cross is picnic white and pink green patch capped by a bear."
Add to this the fact that all neural networks are inherently biased by the information relevent to their creators, and AI•Scry can pop out some absurd results. On this topic, co-founder Tara Shi tells The Creators Project, “The image database that our neural net is trained on, Microsoft COCO, specifically features objects that a 4-year-old could recognize. As a result, it sees a lot of dogs, donuts, bananas, and cell phones in weird places.
While some may argue that this technology is just in its infancy and it will get better over time, no amount of technological development will overcome the social biases instilled in a system by its designers. The machine can only generalize from what its long chain of human trainers know and think about the world. AI•Scry is surprisingly accurate in a typical Western-style kitchen, for example, but would fall to pieces as an intelligence tool in the African Savannah.” The point of the conceptual app is to familiarize people with the technology of artificial neural networks, because they'll likely run the world soon—which might not be a bad thing.
We tested AI•Scry on a bunch of stuff around our office, and here's what it came up with.