Developer Guide for Intel® Data Analytics Acceleration Library 2019 Update 4

Exponential Linear Unit Backward Layer

The forward exponential linear unit layer computes value of the following function



where α is the user-defined real valued coefficient, x is the input of the layer. The backward ELU layer computes the following value



where g is the input gradient tensor obtained from the preceding layer during the back propagation process and z is the value to be propagated to the next layer,

denotes derivative of the forward function



calculated at the point x

Problem Statement

Let

be the input gradient tensor with the components

and let

be the user-defined coefficient. The problem is to compute the tensor

with the components

to be propagated to the next layer



where



for all