I did some experiments that really highlight how heavily image recognition algorithms depend on probability - and how unprepared they are for surrealism. #machinelearning
http://aiweirdness.com/post/171451900302/do-neural-nets-dream-of-electric-sheep
@freundTech Thanks for the link! I hope they'll post some of the results sometime - it would be interesting to read.
@janellecshane You're a star. Please tell me that joking around with neural nets is all you do or that at least you've managed to make a career out of it.
Btw, no advertising for your Masto account on your blog?
@JordiGH I have a day job that has nothing to do with neural nets!
Hmm, I'll have to add my Mastodon link somewhere on my blog. It's on my Twitter bio, but not Tumblr at the moment.
@janellecshane Of course you saw the HN discussion, right?
@JordiGH Didn't see that discussion - thanks for the link!
@janellecshane there was a mastodon bot that "recognized" pics ... Where is it now?
@saper @janellecshane
are you thinking of @fascists_exe ?
@janellecshane @saper
Or possibly @neuralgae , though that also generates an image.
@enkiv2 @janellecshane @neuralgae Nope,
I meant @neural_tv
But it stopped working (@ryanfb) ?
@janellecshane i am now going to greet every sheep i encounter "hello, flower"
@janellecshane Thanks a lot, just what I needed before going to sleep.
My student trained some classifiers on ImageNet and the result was that all quadrupeds were predicted as dog.
We then switched to MS Coco and found that -- while the caption generation is often good for a laugh -- the object detection was not pretty good in most cases.
Not detecting sheep on trees maybe shows that the Deep Networks now have actually a good "common sense"
I wonder how it fare on the columbine harvester picture?
https://www.reddit.com/r/funny/comments/7rkvq4/i_had_to_look_a_few_times/
@deeds Ha, what a picture! As you predicted, Microsoft Azure (trained on Coco, I believe) reported it as "a crowd of people"
@deeds But Cloudsight is an interesting case! It seamlessly uses humans for the hard one and since the caption was part of the photo...
Object: "green and yellow combine harvester"
Scene: "This Look Like A Sick Concert Text"
@janellecshane Surrealism relies upon contextual comprehension.
That is an AI-hard problem, and until we have functionally structured AI designs, that will be very hard to achieve, and effectively impossible to prove if we do achieve it.
Conclusion: the tools we are using to recognise images are still too basic (and we understand them quite poorly).
The pics and the comments are coool 👏
@janellecshane You might find this interesting:
https://open-codes.zkm.de/en/work/rzl-dzl-ai-frederik-de-wilde
It's an art project, which uses AI to generate patterns, which fool other AIs.
Those patterns for example make an AI recognize a tank painted with it as an umbrella or similar.