blob: 7010bd93f2be6bcd01dccf298dab31f999259e5a (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
|
use crate::ctx::Ctx;
use crate::quad::{Node, Embed};
pub fn step<A: Embed, B: Embed>(ctx: &mut Ctx<A, B>, node: Node<A, B>) -> Node<A, B> {
// pad the graph if needed.
// if we're at the base, run the cellular automation rule:
// collect training pair base -> rule.
// return the updated cached node.
// forward predict all children nodes of the root
// collect children nodes into vector X, X'
// forward predict compressed representation of root Y, Y'
// our target is to train Y' to match X'.
// calculate the difference between X and Y.
// if the difference is below an acceptable threshold:
// collect training pair Y -> X' (with error Y').
// return the updated cached node with compr = X'
// if the difference is above an acceptable threshold:
// recurse `step` on each child to produce four vectors
// (the recursion will stop when the true base is reached or threshold is small enough.)
// collect all children predictions into Z'
// collect training pair Y -> Z' (with error Y').
// return the updated cached node with compr = Z'.
todo!("Implement a step!")
}
|