Developer Guide for Intel® Data Analytics Acceleration Library 2018 Update 3

Element-wise Sum Backward Layer

The forward element-wise sum layer computes the value tensor Y as follows:



where are the real valued user-defined coefficients, and the tensor X (i) is the output of the previous i-th layer,

. The backward element-wise sum layer computes the following values:



where G is the input gradient tensor computed on the preceding layer during backpropagation process and Z (i) is the value the layer propagates to the next i-th layer.

Problem Statement

Let be the input gradient tensor with the components and be the predefined scalars. The problem is to compute the tensor with the components to be propagated to the next i-th layer: