In the summer of 2015, Google released DeepDream, a neural network that transformed images into hypnotic hallucinations. The internet went nuts. It was one of the first instances of an experimental project that demonstrated what neural networks were capable of to the public, giving us a visceral glimpse at the future of AI.
At the end of 2016, that future, well, hasn’t quite arrived yet. However, this year we saw AI truly enter mainstream dialogue, as society confronted the sticky ethical implications of its design and regulation.
Meanwhile, alongside this serious debate, we saw a multitude of highly visible, experimental, and sometimes very silly projects borne of AI. As designers and engineers began to play with the technology in a more accessible way, their experiments, while far from groundbreaking, offered insights into where AI is going–along with plenty of entertainment.
Take Google’s A.I. Experiments site, which is a repository for test projects that use machine learning and neural networks to accomplish whimsical, seemingly silly tasks. In one, called Quick, Draw!, you play a version of Pictionary where the Google algorithm tries to guess what you’re drawing while you sketch. Another neural net experiment takes your handwriting and tries to generate more of it. These experiments are fun, but Google is already using some of these same kinds of algorithms to detect objects and faces within Google Photos.
Another real-life challenge that AI could help with? Sorting through images and deciding which are SFW and which are NSFW–a task that would be pretty handy for a search engine to complete automatically. Yahoo has an AI that can divide images into these two categories, but PhD student Gabriel Goh decided to take that algorithm and use it to generate new images that would still fit within the SFW and NSFW buckets. The results? The G-rated images resemble rolling hills, with peaceful streams running through them. But the other batch of images seem like they came from a computer with the mind of a 13-year-old boy–it sees penises, breasts, and vaginas everywhere. “If anything, it shows [Yahoo] is doing its job,” Goh told Co.Design. Good to know it works.
Training AI to generate new images of real, recognizable things was a theme this year. After the Pokémon Go craze hit, one Japanese researcher took the logical next step and built a neural net that conjures new Pokémon (though the results look more like mashed-up impressions of the original monsters). The artist Maja Petrić worked with a Microsoft machine-learning researcher to build a program that would create images of the sky as different people see it around the globe. Meanwhile, a group of Italian engineers inspired by the novelist Italo Calvino’s prose poem Invisible Cities created a neural network that would generate maps of imaginary metropolises.
None of these generated images are even close to realism, of course, but they point toward a future where images of places and things may be created entirely from a computer program.
That power gets quite a bit scarier when it’s applied to people’s faces–literally. One project from Mario Klingemann, Google’s machine-learning artist-in-residence, generates portraits of human faces, and yes, they look like perverted demons come to haunt us. But the project also demonstrates how neural networks process faces; they do so similarly to humans, focusing on the eyes and mouth. “My ultimate goal is to create a machine that can produce ‘art’ entirely autonomously,” Klingemann told Co.Design. “But I’m definitely not there yet, so right now, I’m in the phase where I build tools that produce serendipitous visuals that are different or ‘new’ in some way.”
Over at MIT, a group of engineers decided to put AI to work scaring us–just in time for Halloween. Their project, “Nightmare Machine,” generates images of faces and places that are meant to be terrifying. Their fascination is less in the visual elements of each spooky portrait but in which ones actually spark genuine fear. This research could even inform the next generation of haunted houses: “Maybe in the future, we can generate personalized horror images where we tailor the generation process to the individual data,” Pinar Yanardag, one of the scientists, told Co.Design.
Another project from MIT used AI to tap into human fear of a different sort. Postdoc researcher Bradley Hayes trained a neural network on the tweets of Donald Trump, and then built a Twitter bot–known as DeepDrumpf–that mimics the president-elect. The computer-generated tweets appear uncannily close to the real thing.
We also saw AI being utilized in creative fields, and even the art world, where human taste reigns supreme. At the Tate Britain this year, a computer program called Recognition played the part of curator, dredging up images from the museum’s digital archives that matched photographs of current events and people. The project even got an exhibition of its own to boot.
From conjuring new Pokémon to generating images of penises and mimicking the tweets of Trump, many of these projects are hardly life-changing, but the technology behind them is revolutionary. Taken together, these seemingly inconsequential experiments show us the edges of technology, bringing the great grandparents of the algorithms of the future into focus so we may examine, critique, and understand them.