An activation function in a neural network introduces non-linearity into the model, enabling it to learn complex patterns in the data. It determines the output of a node given an input or set of inputs, which is crucial for the network's ability to approximate any continuous function.
