Hacker News new | past | comments | ask | show | jobs | submit login

One thing that has fascinated me about this is the possibility of translation across dissimilar categories.

What I mean is we have style transfer in the domain of text, and we can style transfer with images. And we can generate images from text. Can we style transfer from image to text or vice versa? Can prose be rewritten in a manner that, in some sense, adheres to aesthetic principles of impressionist painting?

Presumably there would be some kind of informational representation of text style discernable to an image generation system. And just like an artistic style can be extracted from a painting and transposed to a photograph, perhaps an interpretation of textual style could be applied to a photograph despite them being different mediums.

What would that even look like? I don't know, but I find the possibilities fascinating.

The temptation, I think, is to make a first pass at answering this question in a frustrating, cartoonishly shallow way. And I think systems will possibly be developed that just go ahead and do it before people are culturally ready to understand it in a non-frivolous way. Everyone needs to get those reactions out of their system, I guess, but there's a more nuanced possibility here that might allow clashing of dissimilar categories in ways never previously contemplated.




Nice ! This points me back to the my favorite mental model of machine learning/nn. It’s always about shuffling in an out of a number of dimensions and the mappings between them.

When you make this please name it SynesthesAI




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: