Computer-Generated Image Captions
cs.toronto.eduI like how two of the options to this one (http://www.cs.toronto.edu/~nitish/nips2014demo/results/84291...), of a guy at the gym, are
"a woman in a kitchen, leaps in the air while attempting to balance a glass cup in one hand."
and
"a young man playing wii in front of a large knife."
I was going to write a negative post about how singularity is still a long while away based on this caption: http://www.cs.toronto.edu/~nitish/nips2014demo/results/84542...
but then I realized that I cant come up with a description of what is going there myself.
Ill get you next time CPU!
Looks like they are dyeing cloth to me.
I would really like to know how this works because the captions are... interesting.
Because you have never experienced what it is happening in the picture?
The singularity is not going to be an AI revolution, it's when we decode the human brain's signals.
Indian textile workers dye cloth in large outside pots?
Some results are pretty funny.
http://www.cs.toronto.edu/~nitish/nips2014demo/results/84824...
Generated caption: "a man appears to be a banana on a tree"
Are these supposed to be funny?
http://www.cs.toronto.edu/~nitish/nips2014demo/results/92679...
"a man wielding an electric razor is gleefully shaving away another man ' s hair ."
Hilarious!
I think they've stumbled on computer-generated comedy. Some funny stuff in there.
That's from the "Nearest Caption in the Training Dataset". Which means it found the most similar image, and that image had that caption.
No. That is not how it works. Read the papers again.
It very clearly says "Nearest Caption in the Training Dataset". The generated labels are below it.
What papers?
This one http://www.cs.toronto.edu/~nitish/nips2014demo/results/82676... has a tag "fingering"
http://www.cs.toronto.edu/~nitish/nips2014demo/results/80282...
It must be making some very broad generalisations to come up with the tag 'homosexuals'...
It weirded me out, too.
The only image I tried (http://www.cs.toronto.edu/~nitish/nips2014demo/results/79355...) is tagged "lesbianism".
OP, could you give us more details about this ?
Not OP, but from the site... http://www.cs.toronto.edu/~nitish/
Nitish Srivastava, co-instructor for CSC 321 : Intro to Neural Networks ( http://www.cs.toronto.edu/~rgrosse/csc321/ ) pretrains a convolutional neural net using image sets ( https://github.com/torontodeeplearning/convnet/tree/master/e... and https://github.com/torontodeeplearning/convnet/tree/master/e... )
also has a demo to upload your own images and get them captioned or classified http://deeplearning.cs.toronto.edu/i2t but seems their servers are getting blasted right now