Why is there a factor of 1.7159 with the tanh function used in neural network activation?

1.1k Views Asked by At

I was reading about neural networks when I came across the line : Recommended f (x) = 1.7519 tanh (2/3 * x). How do we arrive at these values (we can fix the other once the other is obtained using the condition f(1) = 1) ? Pg 10 at Efficient Backprop

1

There are 1 best solutions below

0
On

If you read further, at the top of page 14 it states that the required conditions for the sigmoid are:

  1. $f(\pm1)=\pm1$

  2. The second derivative is a maximum at $x=1$

  3. The effective gain is close to 1

Once you've decided that a $\tanh$ curve is a useful curve to try to fit to your sigmoid, then it is a case of choosing prameters.