Gradient
Also known as: gradients · gradient descent
A gradient is the direction (and steepness) that tells a model how to adjust each weight to reduce its error. Think of standing on a hilly landscape: the gradient points downhill.
Gradient descent is the process of taking small steps in that direction, over and over, until you reach a valley (a good set of weights).
- Steep gradient → big, confident step.
- Flat gradient → tiny, careful step.
Plain version: gradients are the model's sense of "which way is better". Training is just following that sense, billions of times.