4 ms·
This seems relevant: In the days when Sussman was a novice, Minsky once came to him as he sat hacking at the PDP-6. "What are you doing?", asked M
by vecter 12y ago
This seems relevant:
In the days when Sussman was a novice, Minsky once
came to him as he sat hacking at the PDP-6.
"What are you doing?", asked Minsky.
"I am training a randomly wired neural net to play Tic-tac-toe", Sussman replied.
"Why is the net wired randomly?", asked Minsky.
"I do not want it to have any preconceptions of how to play", Sussman said.
Minsky then shut his eyes.
"Why do you close your eyes?" Sussman asked his teacher.
"So that the room will be empty."
At that moment, Sussman was enlightened.
[0] http://en.wikipedia.org/wiki/Hacker_koan http://en.wikipedia.org/wiki/Hacker_koan
- blinduck 12y agoI don't quite understand this. A person's preconceptions is the collection of his previous knowledge about the subject. What does it mean to have random preconceptions?
- grey-area 12y agoI think the point is that you can't remove external constraints from a situation simply by willing it (make the people in the room disappear, make your AI independent of assumptions). For example the type of neural net chosen, the reinforcement and other parameters, the rules of the game chosen by the creator are all constraints which are not randomly chosen, so randomly choosing the starting state is not necessarily going to remove preconceptions built in to the model, and it is still choosing some starting state, just not an arbitrary one chosen by the creator. I'm not sure it really applies to this article though, which is about a different problem - saving the time spent worrying about unimportant decisions or those where you don't have enough inputs and choosing to take a random path.
- Terr_ 12y agoI think it means that: (A) There are always "preconceptions" present, even if the programmer isn't working on a level of abstraction that lets them consciously realize it. (B) Programming a system to be too naive can backfire.
- sanityinc 12y agoI believe the point is that the neural net will still have preconceptions -- but the random wiring just means you've chosen to not see what they are.