The past couple of years have seen artificial intelligence (AI) improve by leaps and bounds in its mission to steal all our jobs and take over the world. It’s beaten humans in a game of “Go,” it’s won horse bets, and it’s even developed the ability to mirror some of the most complex of human behaviors – such as bizarre, racist Twitter rants.
Another moment was when Google’s AI showed off the “dreams” it was capable of creating. The weird and trippy images it produced were based around the artificial neural network’s ability to recognize images of everyday objects and then later recreate images from a text description of it by using its “memory” of them.
New research published in arXiv has shown just how far this form of artificial intelligence has developed over the past three years. Even within the past months, the progress is noticeable and fairly stunning.
Some of the images the AI has created, using only a text command and the memory of photographs from ImageNet. Image credit: Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, Jeff Clune/arXiv.
Not only does this research have plenty of potential applications in terms of image recognition, the real crux of the work is to give an indication of the power and “deepness” of the neural networks.
The researchers started their paper by saying how their work was inspired by neuroscience’s understanding of the brain and its mechanisms of imagining abstract concepts. The process essentially works by telling the AI what it’s “looking at” and then feeding it images of the object through different layers of artificial neurons. As the information – an image in this case – is processed by each layer, it “notices” different key features and extracts more data. Each layer builds on this decision until the machine can eventually reach an understanding of what an object is meant to look like.
The process can then be flipped around, whereby the AI is able to generate an image just based on a text command. The researchers are able to fulfill this process layer-by-layer, allowing them to see what each artificial neuron is doing and thereby allowing them to be fine-tuned.
Check out the progress they’ve achieved so far (below). The bottom row of images shows AI’s latest abilities, above which are its former attempts since 2013.
This figure shows the progression of a deep neural network’s ability to synthesize recognizable images. Image credit: Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, Jeff Clune/arXiv.
[H/T: Popular Science]