In the “baby bird learning to sing” task it is natural to wonder how the learned task is organized as it is learned and recalled.
It turns out that other people have been thinking about this sort of thing. I have trying to wrap my head around the difference between extending a sequence and forming a similar but longer sequence. Here are some insights:
Chomsky and the platypus
This is what @subutai had to tweet on it:
@Paul_Lamb had these insightful comments in the baby-bird-learns-dadssong working group:
If I understand the experiment correctly, it demonstrates that pooling/chunking of sub-sequences is based on semantic similarities of the elements next to each other in the sequence. The LLL -> RRR (versus LL -> LR -> RR) point fits well with the idea I have for timing. If action “L” is defined and takes, for example, 1 second to complete, then L encoded with timing 2 seconds would unfold as “LL”, 3 seconds as “LLL”, etc. Instead of encoding that there are 3 “L”'s, instead I encode how long I want to perform the “L” action.
LLL -> RRR can then be pooled, and itself repeated. LLLRRR LLLRRR LLLRRR could be encoded as “perform the LLLRRR action for 18 seconds”.
It would solve the repeating sequence problem that HTM currently suffers from.