COMPUTERS DREAM IN LOVELY LAYERS
Computers can dream! They do it via layered neural networks. Google Brain, Google’s AI team has been exploring their capabilities, creating incredible images and gaining new understandings about how machines can both think and create.
“Inceptionism” introduced us to the concept of neural networks by showing how they “dream,” in wild, hallucinogenic images.
Fantastical creatures floating in the clouds. Surreal landscapes dotted with impossible fanciful objects and structures. Mind-bending patterns that pulse with vibrant colors. These are things that computers daydream about. And they are oddly beautiful.
It turns out, computers dream much the same way we do. They reprocess things they’ve seen before in new ways, mixing and extrapolating data to produce novel images and experiences. They imagine.
Recently Google Brain researchers and their colleagues published a new paper on their progress called “The Building Blocks of Interpretability.” It shows a new way of seeing inside the brains of machines.
“Interpretability” is the ability to understand. The Google Brain team built a series of interfaces for looking inside neural networks to see how they make decisions.
As layers of neurons work together, the machine’s understanding evolves. It begins by detecting edges, then it discerns shapes.
It all comes down to Large Scale Deep Neural Networks (LSDNN), complex multilayered digital networks that can process large numbers of inputs and process them on multiple layers, similar to the way human brains work. LSDNNs are used in machine learning and artificial intelligence applications.
Google trained their neural networks by feeding them millions of categorized images of specific subjects. Each network can be trained to recognize a specific category of data. For example, one neural network might be trained to identify trees, while another specializes in dogs.
Neural networks have been used with great success in image recognition and analysis. And when these LSDNNs are asked to recognize the impossible, they hallucinate -- they dream up creative answers, imagining what they might be.
The LSDNNs are comprised of up to 30 layers of networked artificial neurons. Each layer detects different information within the a photograph—from basic edges, to general features, to something as complex as hair length—and puts all that information together to interpret the image.
But then the researchers got tricky. They fooled the machines. They fed photos of clouds into LSDNNs trained to recognize animals, and then asked the networks them to interpret.
Incredibly, the LSDNNs generated images of fantastical hybrid creatures, like pig-snails, camel-birds, and dog-fish. In other words, the computers were able to “imagine” features within the clouds—much like humans do when they daydream and see dragons and castles billowing in the sky.
When humans look for shapes in things like clouds, they imagine that they see familiar things. Artificial machine “brains” do the same thing. Google calls this phenomenon “Inceptionism.”
Things got really interesting when researchers at MIT Computer Science and AI Lab fed the networks random noise and asked them to identify a single feature—like an arch, for instance. Machine imaginations ran wild.
Some of the resulting composites were gorgeous, psychedelic scenes, rippling with disparate objects like odd pagodas, arches, and fountains. Some resembled M.C. Escher’s mind-bending tessellations, only electrified with color.
Deep learning makes for deep dreams. The code that Google used to create its “deep dreams” images is available on Github. Check out #deepdream to see more. If you’re curious, download the code and let you computer’s mind wander!
Read more about Beautiful Layers all this week on BeautifulNow, including Time Travel Via Layered Cities: Vhils and Beautiful Layers: Dustin Yellin Psychogeography, Scrap Layers as Reborn Beauties, and Embroidered Layers Embellished Stories: Melissa Zexter. And check out more beautiful things happening now in BN Wellness, Impact, Nature/Science, Food, Arts/Design, and Travel, Daily Fix posts.
Want more stories like this? Sign up for our weekly BN Newsletter, Like us on Facebook. Follow us on Twitter, Instagram, Pinterest. Join our BeautifulNow Community and connect with the most beautiful things happening in the world right now!
Do you have amazing photos? Enter them in this week’s BN Photo Competition.
