A more robust way to improve the performance of SGD when plateaus are encountered is based on the idea of momentum (analogously to physical momentum). More formally, a momentum is obtained employing the weighted moving average of subsequent gradient estimations instead of the punctual value:
The new vector v(t), contains a component which is based on the past history (and weighted using the parameter μ which is a forgetting factor) and a term referred to the current gradient estimation (multiplied by the learning rate). With this approach, abrupt changes become more difficult, and when the exploration leaves ...
