Backpropagation with Momentum |
---|
This algorithm is based on the basic backpropagation algorithm but introduces a momentum term. This term is used to avoid oscillations around a particular minimum. To the standard equation for computing weight change, we add a term that multiplies the weight change in the previous iteration by a momentum factor. This tries to avoid oscillation problems common with the regular backpropagation algorithm when the error surface has a very narrow minimum. The net effect is that of traversing flat error surfaces quickly, while moving slower when the surface becomes irregular.
|