Developer Guide for Intel® Data Analytics Acceleration Library 2019 Update 5
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

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 