It would be far more terrifying if the rate of change was less rapid and there was a higher degree of continuity between frames.
Right now it's very noisy, presumably because the frames aren't really interdependent. The video description mentions a blending technique, but its primary purpose seems to be ensuring the underlying video is not completely overwritten by the "dream" generation feeding into itself.
Have you or your wife had any big psychedelic experiences? If you haven't and your wife have that's probably the explanation. The Deep Dream animations trigger flashbacks like nothing I've seen before.
I'd like to see this for a network trained for more neutral images like places. See for instance the bottom image of [1]. I imagine it would be cool to continuously zoom and iterate through this image in a video.
This isn't using Google's tech though, they built their own version before Google released any of the code. One of the creators actually started a job at Google right after releasing this.
Ha, I was there for that. Ick. The fun thing is to go full screen, cover up the upper left corner for a couple minutes, then try to guess what it's drawing. I got it right a couple of times.
This isn't exactly Deep Dream's equivalent for video. It takes frames individually (smushing them together), rather than having a unified neural network that takes the entire video at once.
I'm really curious to see how much the original training material affects these images.
Is everyone using the same source? There's a lot of doge in there along with that bird that always pops up. Why are faces so prominent? Is that an artefact of the training data or inherant in the algorithm. I would guess the former.
It was trained on Imagenet. 30-40% of Imagenet is just distinguishing different breeds of dogs. And the remainder is mostly other types of animals and random objects.
It's a combination of the fact that the source video takes place in a club with some flashes/strobe effects going on, and the fact that the algorithm looks at things frame by frame (more or less.) When it flashes it sees more/different details than when it is presented with a dark frame, and it interprets those as a dog.
Interesting. Tough to evaluate quality, very subjective. Still, seems like quite a bit of progress have been made since Hinton's networks dreaming of digits in 2006 - http://www.cs.toronto.edu/~hinton/digits.html
I remember doing something similar a way back, only just using nearest neighbor search on relatively large dataset (1m of human faces with different scales/etc, or sounds in the MP3 voice recordings), rather than passing things through up-and-down the neuro-net. The result was very similar. I wonder if one can get a better result with the dogs as well, just by using nearest neighbor. A good baseline...
(link to youtube version below) https://www.youtube.com/watch?v=oyxSerkkP4o