more robo brain
One of the problems with neural nets is that they have a hard time "remembering" things they've learned before. Since i'm not using neural nets to control robots or do anything productive aside from generate weird images, and I don't give a shit about computational accuracy, I decided that the "hack" solution was to create a long term memory on disc. A neural net may not be able to remember megabytes worth of image data, but it CAN remember the three things needed to access it's "long term" memory, which is a) an image index (which picture to 'look at' for reference, b) the location(x,y) in the image it 'wants' to 'look at' and c) how relevant, and thus, how much influence this input will have on the various neurons in the network.
This is the result of giving it a picture of abe lincoln to "study". I hope you like it!!!!!!
posted by d.j.z. at
10:50 PM
2 Comments
