Developer Guide for Intel® Data Analytics Acceleration Library 2019 Update 2

Dropout Backward Layer

The dropout activation layer applies the transform y = B(r) * x / r to the input data, where B(r) is a Bernoulli random variable with parameter r. For more details, see the forward dropout layer. The backward dropout layer computes value z = B(r) * g / r.

Problem Statement

Given a p-dimensional tensor G = gi1...ip and M = mi1...ipof size n1 x n2 x ... x np, the problem is to compute a p-dimensional tensor Z = (zi1...ip) of size n1 x n2 x ... x np, where:

zi1...ip = mi1...ip * gi1...ip