is it possible to have a learning system generate structured data such as svg or html code? NN are obviously quiet good at generating image data because they can do that on a pixel by pixel basis and small errors in the color tone of one pixel in an image is absolutely acceptable. But how would one go about having a NN generate things like html or svg code? Can it be done? Could a NN learn to generate icons for applications based on a training set of similarly looking (themed) icons to have it generate an icon for that set with the same theme but a different "base"?
Create a learning system that can generate highly structured data such as SVG graphics or HTML trees
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Yes. This is known as structured prediction, and usually done with some sort of recursive neural net. Often for code, there is some inherent tree-structure which can be exploited if your RNN is able to output a tree instead of a sequence.
See pix2code and tree-structured decoding for more details.
