Gradient
The direction and steepness of change that tells each weight which way to move to reduce error.
Think of it like
Like feeling the slope under your feet in fog — it tells you which way is downhill even when you can’t see the valley.
Example
After a wrong prediction, the gradient for each weight says “nudge me up a little, me down a lot,” and the optimizer follows those hints.
How it actually works
A gradient is the derivative of the loss with respect to every parameter — a giant vector pointing toward steepest increase, so training steps the opposite way. Backpropagation is just the efficient method for computing it, and problems like vanishing or exploding gradients are what make deep networks hard to train.
For product teams
The invisible feedback signal that makes learning possible; when it misbehaves, training stalls.
For engineers
∂loss/∂θ for all parameters; computed via backprop and consumed by the optimizer to update weights.
Related
- Learning signal — The corrective signal it carries.
- Training — What consumes it to update the model.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome