Skip to main content

very manual plasticity choices

Created on September 15|Last edited on October 4
Done with 10% of weights having the multiplied(larger) plasticity.

100200300400500600Step11.21.41.61.82
Run set
11

1%

Run set
11

with decay

Run set
9

Call it a "treadmill". More comprehensive hyperparameter sweep involving all three values. The decay rule is now simplified to exclude the norm.

Run set
15

another go at it, something seems fishy.

Run set
50

woah, check out the minimums

Run set
50

benchmark against backprop

Run set
4