Yes it's a bit like a Rorschach inkblot test as used for psychiatric evaluations. In fact I tagged the image 'boy sipping tea' and another LJ user thought it was the 'silhouette of a boy with great hair' :)
It's from picbreeder.or (http://picbreeder.org/) - a genetic art website. The images are generated by applying x,y coordinates (and also a radius) into a neural network inputs and reading the colour for each pixel from output neurons.
As a user you get to pick images that you like most and they are 'bred' to generate a new page of images. The breeding is performed using NEAT (http://www.cs.ucf.edu/~kstanley/neat.html) - that is, the network structures are encoded as a NEAT genome and therefore NEAT crossover and mutation can be used.
Given a network it is possible to tweak connection weights or the network structure or the input values of, say, auxilliary inputs. This allows variations on an image to be generated, e.g. variations on a boy sucking his thumb. This of course could form the basis of a CAPTCHA system with a database of networks each representing some thing (boy, girl, car, plane, etc.) and the CAPTCHA system creating variations on each of these images by tweaking the networks.
It wouldn't always work because sometimes small variations in network inout can result in wildly different images, these networks can be be filtered out automatically with a little thought.
I'm guessing though that a Restricted Boltmann Machine could be trained to associated 'boy' with all of the variations. After all the concept 'boy' exists in a high dimension space represened by an ANN and it's inputs. RBM's have been demonstrated to characterise high level features from sets of images exhibiting variations on a theme/concept, such as a face or handwritten letters and digits.
Interesting idea though to use tagged images instead of warped words. You could even tailor the images to a specific demographic in an attempt to thwart those indian captcha companies. E.g. images of national celebrities not well known outside the US, car models, brand logos (change the words but keep the font/style).
Interesting. Of course $2/1000 captchas is a not insignificant sum of money, so the captcha is still effective in limiting unwanted access to an extent.
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
krja![[identity profile]](https://www.dreamwidth.org/img/silk/identity/openid.png){kind=small}
(no subject)
Date: 2008-09-02 08:21 am (UTC)(no subject)
Date: 2008-09-02 08:25 am (UTC)(no subject)
Date: 2008-09-02 09:38 pm (UTC)It's from picbreeder.or (http://picbreeder.org/) - a genetic art website. The images are generated by applying x,y coordinates (and also a radius) into a neural network inputs and reading the colour for each pixel from output neurons.
As a user you get to pick images that you like most and they are 'bred' to generate a new page of images. The breeding is performed using NEAT (http://www.cs.ucf.edu/~kstanley/neat.html) - that is, the network structures are encoded as a NEAT genome and therefore NEAT crossover and mutation can be used.
Given a network it is possible to tweak connection weights or the network structure or the input values of, say, auxilliary inputs. This allows variations on an image to be generated, e.g. variations on a boy sucking his thumb. This of course could form the basis of a CAPTCHA system with a database of networks each representing some thing (boy, girl, car, plane, etc.) and the CAPTCHA system creating variations on each of these images by tweaking the networks.
It wouldn't always work because sometimes small variations in network inout can result in wildly different images, these networks can be be filtered out automatically with a little thought.
I'm guessing though that a Restricted Boltmann Machine could be trained to associated 'boy' with all of the variations. After all the concept 'boy' exists in a high dimension space represened by an ANN and it's inputs. RBM's have been demonstrated to characterise high level features from sets of images exhibiting variations on a theme/concept, such as a face or handwritten letters and digits.
Interesting idea though to use tagged images instead of warped words. You could even tailor the images to a specific demographic in an attempt to thwart those indian captcha companies. E.g. images of national celebrities not well known outside the US, car models, brand logos (change the words but keep the font/style).
(no subject)
Date: 2008-09-02 02:47 pm (UTC)(no subject)
Date: 2008-09-02 09:12 pm (UTC)(no subject)
Date: 2008-09-02 11:20 pm (UTC)(no subject)
Date: 2008-09-03 05:24 am (UTC)