⟵ Back to Course Overview

Activation Functions Visualizer

Non-linearities and Gradient Flow
Task Description: Non-linear activation functions are essential in obtaining expressive functions with neural networks. Some of these functions are better than others, and in most cases it is due to what they do to the gradients during backpropagation. Take a look at these functions, and if they have any adjustable parameters see the influence on the function and derivatives.
Function $f(x)$ Derivative $f'(x)$

Vanishing Gradients: Watch the blue dashed line. When it approaches zero, the weight updates in early layers of a deep network will effectively stall.