April 2020
Intermediate to advanced
330 pages
7h 44m
English
One potential problem with ReLU is known as dying ReLU, where, since the function assigns a zero value for all negative values, signals can get dropped completely before reaching the output node. One way to try to solve this issue is to use Leaky ReLU, which assigns a small alpha value when numbers are negative so that the signal is not completely lost. Once this constant is applied, the values that would otherwise have been zero now have a small slope. This keeps the neuron from being fully deactivated so that information can still be passed on to improve the model.
Let's create a simple example of the Leaky ReLU activation function:
Read now
Unlock full access