One direction for every weight

Take one weight with a current value of 0.60.

Backpropagation asks two small questions. If this weight moved slightly up, what would happen to the error? If it moved slightly down, what would happen?

The answer for one weight is part of the .

A positive value means increasing that weight would increase the error nearby, so training should move the weight in the opposite direction. A negative value points the other way. A value near zero says that a small change would make little difference right now.

The gradient points toward the steepest local increase in error. To reduce the error, training steps in the opposite direction. The weights have not moved yet.

One direction for every weight

Take one weight with a current value of 0.60.

Backpropagation asks two small questions. If this weight moved slightly up, what would happen to the error? If it moved slightly down, what would happen?

The answer for one weight is part of the .

The gradient points toward the steepest local increase in error. To reduce the error, training steps in the opposite direction. The weights have not moved yet.