The Guardian is reporting that Google is trying to understand how its neural net for image recognition works by feeding in random noise then telling the neural net to look for certain features then feeding the resulting image back in. Apart from anything else some of the images generated are astounding.
Link to original Google research article.
(Score: 2) by fritsd on Sunday June 21 2015, @12:52PM
I found it awesome as well, I'm going to try it out if I ever find the energy.
I vaguely remember how to program a backprop, and I know a little bit about the problem that an m → n dimensional mapping, with m >> n, cannot be inverted. Besides, with a sigmoid activation function, the mapping isn't even linear.
So how do you construct an (m-n) dimensional basis set to "espalier" [bbc.com] the missing dimensions?
Backprop is pretty simple to understand, maybe what these researchers found is simple as well.
The images remind me of Jeroen Bosch's painting "de Tuin der Lusten [wikipedia.org]" (might be NSFW or make you crazy). Or like when you read Mythago Wood [wikipedia.org]. Quite disturbing, really.
(Score: 2) by acid andy on Sunday June 21 2015, @07:27PM
These could be self associative neural networks e.g. Hopfield networks. I am not sure if back prop is used with those, though I'm no expert.
If a cat has kittens, does a rat have rittens, a bat bittens and a mat mittens?
(Score: 0) by Anonymous Coward on Monday June 22 2015, @03:52AM
Their algorithm sucks if it is supposed to recognize bananas but does so even if there are no bananas... I think this is just spin on a failed project.
(Score: 2) by fritsd on Monday June 22 2015, @10:55AM
Their algorithm sucks if it is supposed to recognize bananas but does so even if there are no bananas... I think this is just spin on a failed project.
Failed projects are sometimes the most interesting projects. Have you never heard about serendipity [wikipedia.org]? Like how Teflon [wikipedia.org] was invented? Tefal frying pans FTW.
(Score: 0) by Anonymous Coward on Monday June 22 2015, @07:19PM
No disagreement there. I just disagree with (the possible) obfuscation of the motivations for the project.
(Score: 2) by jasassin on Wednesday July 01 2015, @06:53AM
Wow. Some far out artwork. Thanks for the link! Very interesting.
jasassin@gmail.com GPG Key ID: 0xE6462C68A9A3DB5A