Google's Deep Dream Artificial Neural Network

The Something Awful Forums > Archives > General Bullshit > Google's Deep Dream Artificial Neural Network

zedprime: Jun 9, 2007; yospos

gnarlyhotep posted:

they made a program that puts animal faces over human faces

truly astounding

Its not just animal faces, you could make it put dicks all over pictures too.

gnarlyhotep posted:

On the contrary, I think they will eventually hit on something good. They already have done a lot of really great stuff, the street-level mapping of most of the civilized world is amazing. This on the other hand is something that a college grad would do for recognition on his CV.

Sabel beat me to it

Its the stress release portion of an actual useful tool they use to judge reverse image search algorithm and data-bank robustness.

# ¿ Jul 14, 2015 03:27

Adbot: ADBOT LOVES YOU

# ¿ Apr 27, 2024 21:56

zedprime: Jun 9, 2007; yospos

Al Cowens posted:

ANNs while interesting in their own right are not particularly new and it's not all obvious what new spin or advance they have made in an approach that several decades old besides throwing a lot of compute cycles at the problem and the results are hardly visually compelling (subjective, opinions are like assholes) and what we do see probably would be better off just slapping a few photoshop filters on a per-frame basis or a pixel shaders to get something that can be done on any single mid-range desktop computer of 201x.

Also I have never been on LSD or hallucinated so I guess now I have a better idea how what that looks like beep boop beep boop is this what it's like to be human

e: What I see when mom wakes me up for school

Except for some flowery language about the neural net dreaming in the accompanying blog post, the Google devs aren't really trying to pass it off it off as something revolutionary. It started as a way to judge the potential for false negatives on the neural net they were looking at for image crawling: assembling a test set with unique examples is kind of counter intuitive since you presumably want to feed it every unique example you have to begin with and the raw data banks of a neural net aren't exactly interpretable. So they set up the feedback mechanism as a way to see what the neural network has come to learn what the subject looks like to subjectively judge the learning algorithm and the input data set.

Primary use case usefulness aside they thought it looked pretty cool and massaged it into a public release thinking maybe some nerds will be into this, and they were apparently right because the internet has blown up about it.

E. Putting the pieces together it looks like the internet blew up first: the initial blog post was just an example of neural net capability analysis and then internet nerds clambered for the source because they liked the pictures so much.

zedprime fucked around with this message at 12:49 on Jul 14, 2015

# ¿ Jul 14, 2015 12:43

zedprime: Jun 9, 2007; yospos

Anyway when I was 16 I made two Markov string generators talk to each other and I won't be convinced that that, as well as this, is not art.

# ¿ Jul 14, 2015 13:00

zedprime: Jun 9, 2007; yospos

I have to wonder if this would have made waves beyond a more specific computer science crowd that originally asked for the release if the reference implementation didn't put eyeballs everywhere.

# ¿ Jul 15, 2015 18:42

The Something Awful Forums > Archives > General Bullshit > Google's Deep Dream Artificial Neural Network