What functions does something like this? Thanks in advance
>>8194131
https://en.wikipedia.org/wiki/Sigmoid_function
>>8194134
l need that the steepnest varies in function of a like pic related, the smallest a is the more step the function is and the bigger a is the less step the function is
That looks like an arctangent function compressed to a range of (0,1)
>>8194131
You're probably talking about the integral of a normal distribution gaussian with the width given by a. As far as I know it doesn't have a neat algebraic form.
Note that as a goes to 0 the gaussian becomes a dirac delta and the integral of a dirac delta is a step function which is what >>8194141 is probably talking about
>>8194141
1/(1+e^(ax))
the higher the absolute value of (a) the more "step-like" it is.
>>8194195
This. Something like
CDF[NormalDistribution[a/2, a/5], x]
should do it (in Mathematica, plug that into Wolfram Alpha with your desired "a" value to get your function).
>>8194131
logistic function
Seriously guys? It's obviously tanh
>>8195771
Well, you're right, but it is compressed and translated.
>>8195771
It's any number of things you fool.
>>8195778
I agree with you more than that guy. That said, tanh is a good choice.
>>8195778
Quantities are a social construct.
1/(1+exp(k*(x + a))) for some positives k, and a, I think
>>8194141
you are trying to decrease the learning rate wrt the cost?
i use >>8195915
just shift the sigmoid -4 and squeeze by a factor of 4 (or 8 but then you should divide the learning rate by 2 as you don't really want to be over 0.5 in most cases)
also i don't know why this isn't the norm, been doing it for a good year now, momentum is a retarded solution for this.
>>8194131
arctan(x)
>>8195973
*arctan(ax)*2/pi
Look up logistic curves
>>8194131
[math]f(x) = - \frac{\textup{erf}(\textup{ln}(\frac{a}{x} - 1)) + 1}{2}[/math]
http://www.wolframalpha.com/input/?i=plot+(erf(-ln(2%2Fx-1))%2B1)%2F2+where+0%3Cx%3C2
Sigmoid functions have domain (-∞,∞) and range (-1,1), so even if they have a similar contour, they're not valid answers. You have to first map (0,a) to (-∞,∞), which you can do by taking the inverse of another sigmoid function (here the logistic function) with appropriate shifting.
>>8195987
this
looks like a distribution function in [0,a]
>>8194131
>Graph something to an odd power
> turn paper 90 degrees
>>8197365
polynomials don't have asymptotes, genius
>>8197374
I didn't say a polynomial asstard.
>f of x to the third
das da the hemoglobin binding curve senpai
>>8197379
x to the third is a polynomial, it won't have asymptotes, and if it's a monomial of odd degree then it's a diffeomorphism. so unless f had asymptotes in the first place (it was already the function you wanted) then composing it with cube is not going to do shit
retarded fucking newfag
>>8194131
The error function or complementary errorfunction.
Scale it and shift it.