20070911

Prototype

The first image from an effort to merge perceptron and electricsheep


7 comments:

  1. Looks really nice. I love the fine structure you can get with the logarithmic rendering. Towards the middle it's kind of washed out though. Maybe some kind of gamma adjustment or level shift would fix that?

    Out of curiosity, how much processing power do you have left over after implementing this?

    ReplyDelete
  2. Oh, and I hate to see nice images like this marred by JPEG artifacts -- maybe use PNG next time?

    ReplyDelete
  3. oh yes, I see now... I assumed mspaint would default to zero compression, my mistake.

    as for processing power, none. Not only is the written in Java, but its written in a profoundly stupid way. You see, to patch in multiple maps I just replaced the map with a complex vector valued function, as a result lots of little arrays are clogging things up.

    If I took another day to be more clever with the code I could probably free enough cycles at 600x400 resolution to put the rest of perceptron back in, but I think this application really exceeds the limits of Java.

    On another note, it would be really nice if we take advantage of parallel vector computation when processing several maps for each pixel

    ReplyDelete
  4. So you're processing multiple maps per pixel? What do you do? Average the resulting energy densities?

    ReplyDelete
  5. yes, currently its doing a simple average over all maps

    This is possibly less than ideal ?

    ReplyDelete
  6. I thing we can get around the nastyness of branch cuts in the energy accounting by using an algebraic notion of the derivative, instead of numeric. This should be made easier if we are using all functions in their taylor expansions.

    ReplyDelete
  7. also, the rendering is washed out in the middle because the maps are not "contractile" ( I think is the term ) in that region : Any information gets "expanded" and washed out near the origin.

    ReplyDelete