Wednesday, September 14, 2016

ArtInspired :: Deeper into Style & Inceptionism

Inceptionism is the made up name by some of the google software engineers working on a project (deep dream) where they were looking to training artificial neural networks to recognize objects within images It is an attempt the characterize a technique they developed to explain what was happening in their networks. The article in the link above explains some of the process in more detail but I like to think of it as a slightly different way of seeing the knowledge the network is building played back as a superimposed patterns that reflect what they are looking for. The deeper you go the more things look like what you seek and the patterns of line form and light get distorted or re-seen as something else. Positive reinforcement good wild in a meta-self-referencing sort of way. If you are looking for animals, birds or insects, mystical beasts start to appear every where as if a dream (hence the project name) in an artificially intelligent neural network cycling through its meals of data in a computer somewhere in “the cloud”.
#226 Suburban Bush Dreamtime
The deep dreaming code and many of their networks are publically released and I am still keen to code it up and build something for myself one day (probably not soon). However the meme of deep dreams has spread quickly on the internet and their are now many website that offer the deep dream approach. One of the first was dreamscope, and it is my favorite, but google also has an on-line deep dream generator. Many of theses site like lucid for android phones offer a instragram-like pick and one-click filter approach. Already there are already two variants.
  1. Convolutional Neural Networks, these have been describe by a team at University of Tubingen. These neural networks are base on a feed forward approach and they are starting to be able to detect style and the nature of marks made as well as the content. The Tubingen study suggest that style and content are separable.  Thus you can apply a particular style to a  different subject.These system using this approach often have style in their name, (eg google style) and I will adopt that name to characterize this group.
  2. Inceptionist, neural networks. This is the approach described in the google paper (link at the top of this post. It is really just a way to describe googles approach which are extended (more layers) convolution neural networks but have been put into the feedback loop, desperately looking for what they know how to find.
So how useful are these techniques to help an artist create original art?

1. Style

This group is likely to be the most over-used (instagramed to death) because they are in general so easy to use. I am experimenting with such a system called Lucid on my android phone, there are plenty of others for android & apple and even Faux look-alike systems. The better ones like Lucid, do the processing in the cloud, so you load the photo you wait a little while then pick the thumbnail showing the result you want (Lucid uses famous art works for it network training). You pick one and wait a little longer and presto you have something like your photo but painted or sketched by that artist. Some images work, and work wonderfully a lot look just sad. It will be the posting of these unfortunate images that might give this approach a bad name.
A photo posted by norm hanson (@apimageo) on

Hang on I’ve already posted a few myself!
There are some web based “filter” or “app sites” that also offer this style approach. My example here is using google’s deep style approach and a Henri Rousseau painting as my network training example. The results are clearly Rousseau inspired and yet also clearly recognizably the subjects of my photos.
Deep Style using RousseauDeep Style using Rousseau 2
Whilst I have seen many wonderful examples, there are a lot of very trashy examples as well and this my lead to the same reaction of many non photographers to the more extreme HDR works. They may reject these works as a gimmick, without seeing the potential

2. Inceptionism

It is not only google who offer access to the original inceptionism technique, here I am using Dreamscope again but with two different neural training sets following the  feedback approach. The first is called Angel Hair is more about texture and the flow of lines, the second called inception painting is about seeing animals and generates those weird animal eyes and patterns.  You may need to click on the images below to see this detail.
Dreamscope using Angel HairDreamscope using inseption painting
Inceptionism is very quirky, and needs to be used with constraint. Also the deep galleries’ sites in general are becoming far to overloaded with selfies & cats.

Still I see a lot of opportunity to build original art with these techniques.

No comments: