Versioned name: Selu-1

Category: Activation function

Short description: Selu is a scaled exponential linear unit element-wise activation function.

Detailed Description

Selu operation is introduced in this article, as activation function for self-normalizing neural networks (SNNs).

Selu performs element-wise activation function on a given input tensor data, based on the following mathematical formula:

\[\begin{split}Selu(x) = \lambda \left\{\begin{array}{r} x \quad \mbox{if } x > 0 \\ \alpha(e^{x} - 1) \quad \mbox{if } x \le 0 \end{array}\right.\end{split}\]

where α and λ correspond to inputs alpha and lambda respectively.

Another mathematical representation that may be found in other references:

\[Selu(x) = \lambda\cdot\big(\max(0, x) + \min(0, \alpha(e^{x}-1))\big)\]

Attributes: Selu operation has no attributes.


  • 1: data. A tensor of type T and arbitrary shape. Required.

  • 2: alpha. 1D tensor with one element of type T. Required.

  • 3: lambda. 1D tensor with one element of type T. Required.


  • 1: The result of element-wise Selu function applied to data input tensor. A tensor of type T and the same shape as data input tensor.


  • T: arbitrary supported floating-point type.


 <layer ... type="Selu">
         <port id="0">
         <port id="1">
         <port id="2">
         <port id="3">