AI Terminology Course
AI Terminology
/
Advanced

Activation Functions

Definition

Mathematical equations attached to each neuron in a network that determine whether the neuron should 'fire' (activate) or remain dormant, introducing non-linearity to the model.

Explain Like I'm New

A neuron does a bunch of math and gets a number, like 452. The Activation Function acts as a gatekeeper. It looks at the number and says: 'Anything greater than 0 becomes a 1. Anything less than 0 becomes a 0.' It decides if the signal is strong enough to pass to the next layer.

Real World Example

The ReLU (Rectified Linear Unit) activation function is used in almost all modern AI. Its math is incredibly simple: `if (x < 0) return 0; else return x;`. This tiny rule allows the network to learn incredibly complex shapes and patterns.

Common Use Cases

  • •Adding non-linearity
  • •Network design

Interview Questions

basic

  • What happens to a Neural Network if you do not use any Activation Functions at all?

intermediate

  • What does the 'Softmax' activation function do at the very end of an output layer?

Flash Cards

Question

No activation functions?

Click to reveal answer
Answer

The entire 100-layer Neural Network mathematically collapses into a single straight line (a simple linear regression). It would be completely incapable of learning complex patterns like recognizing a human face.

Question

What is Softmax?

Click to reveal answer
Answer

It takes an array of random numbers outputted by the network and squashes them down so that they all add up perfectly to 1.0 (100%). This turns the raw math into readable percentages (e.g., '90% chance it's a dog, 10% chance it's a cat').