/sci/ - Science & Math » Thread #12012187

108KiB, 1366x768, aa.png

View Same Google iqdb SauceNAO

Anonymous Sat 15 Aug 21:22:21 2020 No.12012187 View Reply Original Report

Quoted By: >>12013138 >>12013614 >>12014184 >>12015246 >>12015275

arcsinh is the best activation function

Anonymous

Anonymous Sun 16 Aug 2020 01:11:57 No.12012958 Report

Quoted By:

Yes.

Anonymous

Anonymous Sun 16 Aug 2020 02:09:08 No.12013138 Report

Quoted By: >>12013290 >>12013302

>>12012187
How is it any better than arctan

Anonymous

Anonymous Sun 16 Aug 2020 03:04:03 No.12013290 Report

Quoted By:

>>12013138
It has large gradients

Anonymous

Anonymous Sun 16 Aug 2020 03:06:47 No.12013302 Report

Quoted By: >>12014180

>>12013138
it has large gradients, arctan's get smaller with each layer

Anonymous

Anonymous Sun 16 Aug 2020 05:31:38 No.12013614 Report

Quoted By: >>12013761 >>12014190

>>12012187
>arcsinh
This is nigger notation. It represents area, so it should really just be arsinh. There's no arc involved here.

Anonymous

Anonymous Sun 16 Aug 2020 06:28:38 No.12013761 Report

Quoted By:

>>12013614
how about areasinh or asinh?

Anonymous

Anonymous Sun 16 Aug 2020 09:44:54 No.12014180 Report

Quoted By: >>12014516

>>12013302
That strongly depends on what you want to do with it. For large z, d/dz asinh goes to 0 just as much as tanh or sigmoid, albeit more slowly.

Anonymous

Anonymous Sun 16 Aug 2020 09:49:16 No.12014184 Report

Quoted By: >>12014516

>>12012187
that's a funny way of spelling softsign.

Anonymous

Anonymous Sun 16 Aug 2020 09:52:23 No.12014190 Report

Quoted By:

>>12013614
it's the hyperbolic equivalent of the arcsine, literally arsinh(x) = arcsin(i*x). Stop getting huffy about nonfunctions that pass through infinity.

Anonymous

Anonymous Sun 16 Aug 2020 12:36:17 No.12014516 Report

Quoted By: >>12014702

>>12014184
It's not softsign. It has infinite range

>>12014180
It varies logarithmically. Not like sigmoid

Anonymous

Anonymous Sun 16 Aug 2020 14:00:41 No.12014702 Report

Quoted By: >>12014707 >>12014713

>>12014516
>It varies logarithmically. Not like sigmoid
Do know what derivatives are and why they matter in NNs?

Anonymous

Anonymous Sun 16 Aug 2020 14:02:49 No.12014707 Report

Quoted By:

>>12014702
Yep. The derivatives are different.

Anonymous

Anonymous Sun 16 Aug 2020 14:05:01 No.12014713 Report

Quoted By:

>>12014702
Yeah. the logistic function is limited by an asymptote, log isn't.

Anonymous

View Same Google iqdb SauceNAO based erf.png, 9KiB, 510x331

Anonymous Sun 16 Aug 2020 18:34:05 No.12015246 Report

Quoted By:

>>12012187
yeah i dunno chief

Anonymous

Anonymous Sun 16 Aug 2020 18:42:01 No.12015275 Report

Quoted By:

>>12012187
>activation function

Do people only care about function for neural networks?

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches