Skip to content
BENIBIX Learn

Gradient

General🔢 Math & StatisticsCalculus

Also known as Gradient Vector

A list of slopes, one for each input. Together they point the steepest way up.

Think of it like

It is like an arrow drawn on a hill. It points the steepest way up.

Example

Lucy stands on a hill. The ground climbs 3 metres a step going north and 1 going east. Her gradient is 3 north and 1 east, so uphill is mostly north.

Why it matters

Training an AI model means walking down a gradient. Each step moves the model's numbers so its errors shrink.

Definition

A gradient is the slope of something that depends on more than one number. It is a list of slopes, one for each number. Together they form an arrow, which is a vector. The arrow points the way that climbs fastest. Its length says how steep that climb is. To go down as fast as you can, you walk the other way.
The gradient points straight up the hill. To get lower fastest, step the opposite way.

Common mistake

Thinking the gradient points downhill. It points up, so a model that wants less error steps the other way.

Key takeaways

  1. 1
    One slope per input. Each number gets its own.
  2. 2
    Points uphill. The steepest climb from here.
  3. 3
    Length is steepness. A longer arrow means a steeper hill.