ELU#
Exponential Linear Units are a type of rectifier that soften the transition from non-activated to activated using the exponential function. As such, ELU produces smoother gradients than the piecewise linear ReLU function.
\[
{\displaystyle ELU = {\begin{cases}\alpha \left(e^{x}-1\right)&{\text{if }}x\leq 0\\x&{\text{if }}x>0\end{cases}}}
\]
Parameters#
| # | Name | Default | Type | Description |
|---|---|---|---|---|
| 1 | alpha | 1.0 | float | The value at which leakage will begin to saturate. Ex. alpha = 1.0 means that the output will never be less than -1.0 when inactivated. |
Example#
use Rubix\ML\NeuralNet\ActivationFunctions\ELU;
$activationFunction = new ELU(2.5);
References#
-
D. A. Clevert et al. (2016). Fast and Accurate Deep Network Learning by Exponential Linear Units. ↩
Last update:
2021-03-03