If you looked closely, a pair of eyes had even broken though in the pinkish lower jaw. Overall, it appeared that some horrid infection was brewing underneath the animal’s fur, with teeming sets of snouts and eyes straining to burst through at the next instant. In fact, pieces of dog face popped up in all sorts of unexpected places. Bulging from his canine haunches was a separate snout with another pair of unsettlingly alert eyes. (Maybe not so surprising, given that the network was trained largely on dog breeds.) On the beast’s forehead was a second set of eyes. This is startling to begin with, because the source of this image was not the beagle but the adorable little kitty cat. The image is of a dog, in the broadest sense. With repeated passes of modified images, he got a final output that was not at all normal. His code tapped the neurons mid-process, building the half-baked clues of dogness into more fully realized dogs. But Mordvintsev was hoping for something novel and unexpected. (He found it on a digital wallpaper website.) Normally, one would use a vision-recognition neural net to identify what it saw. He fed a photo into it: a beagle and kitten, each perched on tree stumps with a meadow in the background. The open-source tool he was using to build his neural net had been “trained” on a well-known dataset called ImageNet to recognize objects of 1,000 categories, including 118 dog breeds. But in this pass, Mordvintsev got the balance just right. The actual chunk of computer code that turns a neural net into something that churns out images from its hitherto secret life turns out to be only about thirty lines of code. “It’s easy to write the code and tricky to find the right parameters,” says Mordvintsev. The trick was getting the system to do its thing - reversing itself and then reaching back into itself to find templates for new images - at just the right time and in just the right measures. You could almost taste the pixels being spit out like greasy gravel when the wheels spun on digital asphalt, as the system seized on hints of objects and recklessly took license to flesh them out into vivid representations of a target image. It was like gunning the system forward, then suddenly slamming on the brakes and reversing. Mordvintsev’s process was more Fast and Furious. Previously, the mission of convolutional neural nets was to proceed in a defensive-driving fashion, straining to filter out wrong turns and make accurate predictions. In the middle of the network’s usual practice of trying to verify a nascent sense that a particular pattern may be a target object, he told the network to skip directly to “Go,” and then start making the object. In other words he would flip the function of the neural net from recognizing what was there to generating stuff that might not be there. Just like, whatever it sees in this batch of images, let’s have more of it. Let’s find something that increases the magnitude of the activation vector, he told himself. On this restless night in May, while his wife and child slept, he did the coding equivalent of fiddling the dials to change the objective of the neural net. Mordvintsev wanted to continue down that path, with a wicked turn: He was writing code to make a neural net create meaningful images that weren’t there at all, at least not as humans could tell - visions born of machines, oozing out of the metaphorically neural connections in the system. By looking at those images, the researchers had a better idea of what the neural network was up to at that instant. One team in particular, from the Visual Geometry Group at the University of Oxford, had taken an interesting approach to analyzing how successful vision systems can recognize (classify) objects: at a certain point in the training process, they got the networks to generate images of what they were perceiving. ConvNets are a specialized form generally used for vision recognition they take the biological metaphor farther by not only using a neuron-style learning system, but by employing the neurons in a similar fashion to the way light receptors are arranged in the visual cortex. His curiosity was piqued by one the abiding mysteries of neural nets and deep learning: why did they work so well and what the hell went on inside them? Others had been asking the same question, using what are known as convolutional neural nets (ConvNets) to probe vision recognition systems at various points in the process. As an NN newbie, Mordvintsev was teaching himself about the field, absorbing key papers and playing with systems already trained to recognize certain objects.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |