How to systematically construct sigmoids having wobbly differentials?

43 Views Asked by At

In machine learning using neural networks, sigmoid functions are often used to mimic the dynamics of biological neuron's plasticity. Often sigmoids with a smooth curve is desired, the properties strictly demanded on these functions are usually:

$$f(x+x_0) > f(x), \forall x \in \mathbb R\\\lim_{x\to -\infty} f(x) = -1\\\lim_{x\to \infty} f(x) = 1$$

However as long as we have $f' = \frac{\partial f}{\partial x} \geq 0, \forall x\in \mathbb R$ this should still be fulfilled.

So assuming we have a pair $f, f'$ fulfilling the above constraints, are there any general methods we can apply to periodically (over $t$) "slow down" the learning,


Own work

So far, I have been thinking of, for example something like:

$${f_1}'(t) = (\sin(kx)^{2n}+\epsilon){f_1}'(t)\cdot$$

It is obvious to me that it will satisfy requirements above, maybe with some renormalization of the integral of $f'$ requrired. But in general it seems like a difficult modification because of the difficulty of doing analytic integration. Is there maybe some smarter modification we can do to a general $f$ or $f'$ which always gives an $f'$ for which $f$ is easy to calculate and always satisfies the conditions above?

1

There are 1 best solutions below

1
On BEST ANSWER

Without getting to differential equations, you could try with something like \begin{equation} f(x)=\sigma(k(wx+\sin(wx))) \end{equation} where $k,w$ are constants and $\sigma$ is your "non-wobbly" sigmoid.

I think there is no "correct" answer as the function you are searching for has a too broad definition...