Slashdot Mirror
Turning Neural Networks Upside Down Produces Psychedelic Visuals
cjellibebi writes: Neural networks that were designed to recognize images also hold some interesting capabilities for generating them. If you run them backwards, they turn out to be capable of enhancing existing images to resemble the images they were meant to try and recognize. The results are pretty trippy. A Google Research blog post explains the research in great detail. There are pictures, and even a video. The Guardian has a digested article for the less tech-savvy.
That makes an argument as to what psychedelics actually do.
Any possibility that they will release higher-res versions of these images? Maybe sell some prints?
I realize these are just the output of a funnel-run-backwards, but they'd make awfully cool posters.
An internal system operation returned the error "The operation completed successfully.".
There is a guy that wrote wrote an essay some years ago that suggested as much. He posited that drugs like psilocibin basically overload the brain and cause it to form feedback loops. Many of the effects you can experience on hallucinogens also suggest as much. Closed eye visuals for instance are basically the "lights" you see when you push on your eye balls. They are just amplified and put into a feedback loop. Thought loops are common on hallucinogens as well, I imagine its the result of that as well.
I have seen things very similar to these on some level. My pattern-recognition algorithms aren't fixated on dogs, but the way some of these images look is very familiar. Translucent objects popping out of feedback-looped noise is real. Fractal-repeated infinite patterns are real. Sky turning into crinkly paint swirls is real. Empty space around objects (like the Google logo) showing iridescent patterns is real. Now, there is a caveat that each experience I've had was as different from the previous one as it was from the default world. I also know people whose experiences are vastly different from mine, or even completely non-visual. Not all brains are wired the same!
To some extent yes, but it's likely way more complicated than that. But, yeah, without sensory input we start hallucinating. It's like asking your senses repeatedly "is that real" and the senses always say yes, rather than no. So you drift off into whatever because that's real and therefore there's this other things too.
It's a bit like that old parlor game where you tell somebody that you're going to have them ask questions about a dream, send them in the other room, asking for dream volunteers, and then tell the people still in the room that the answer is yes if the last letter of the last word of the question ends A-M and no if the last letter of the last word of the question ends N-Z. -- They inevitably guess a dream involving all manner of perverted stuff as the crowd confirms and rejects bits at random. Inventing a dream out of his own head rather than somebody else's head.
There's also every day hallucinations like seeing detail where it doesn't exist, movement where it doesn't exist, and hallucinating something to fill the big blind spot in our eyes.
It is no longer uncommon to be uncommon.
It's true. And we know the physical texture does extend fractal-wise into infinity... I'm thinking the opposite is when one is not on psychedelics and is further stressed out, texture details disappear if they are not relevant for the stressful situation. (E.g a sponge becomes a yellow block, no holes or pores.) As if psychedelics open the valve and stress closes it, like many people have said.
I know nothing about these NNs, but the NNs used for the ImageNet Competition typically have a few hundred thousand neurons. This is to place images of about 1M pixels into one of 1000 categories. Most image recognition NNs are "convolutional" which means they are tiled. So each neuron in the first layer is only looking at a small part of the image. Later layers will pool the results from the convolutional layer into extended features. This cuts way down on both the size, and the amount of computation needed. The number of layers will typically be 5 or 6. Even more layers should, in theory, help, but deeper networks are very hard to train. The total size in bytes would be maybe 100MB, but that will vary widely depending on the implementation. I don't know how big Baidu's implementation was (they were the winner, but they cheated). NNs can run fast, categorizing hundreds or thousands of images per second. But it can take a long time to train them, days or weeks on a GPU farm. Fortunately the training is highly parallel.