Neural Networks (fundamentals)

Why each neural network needs an activation function?

Neural networks without activation functions (pure linear combinations of inputs) can only solve linear problems. This is shown in this intuitive and simple example of a neural network incl. some math. → video or here

However, modern neural networks use activation functions specifically to introduce non-linearity, which allows them to approximate any continuous function (this is known as the Universal Approximation Theorem).

What is an activation function?

Relu vs. Sigmoid

Sigmoid

ReLu

Adam