Conversation

Notices

1. Bob Mottram (bob@social.freedombone.net)'s status on Monday, 13-Nov-2017 07:13:50 EST Bob Mottram
   This all sounds great until you realize that it was done multiple decades ago, and it's actually much better to do this via genetic programming because you can generate real working programs with valid syntax in your chosen language, rather than just a bunch of weights.
   
   Unfortunately, and despite some pretty good efforts, automating the task of programming turned out to be really hard even if you have a lot of training data and CPU power. In practice there are not many types of problem which can be programmed for in this manner.
   
   https://medium.com/@karpathy/software-2-0-a64152b37c35
   1. Hattie Cat (hattiecat@quitter.no)'s status on Monday, 13-Nov-2017 15:48:49 EST Hattie Cat

      in reply to
      @bob This neural netty stuff sounds great on one level but you can't reason about it, have no idea how it works (comes up with the answers it does) or know how it'll perform with unexpected input outside of its training set. May not be an issue for some applications I guess.