動量（Momentum）

本節探討權重調節的變化。

This section explores the variations of theweight adjustment.

到目前為止，權重調節依賴於最簡單形式的方程2.7和3.7。

So far, the weight adjustment has relied onthe simplest forms of Equations 2.7 and 3.7.

然而，有各種各樣的權重調節形式可用。

However, there are various weightadjustment forms available.

使用先進的權重調節公式的好處是能夠在神經網路訓練過程中獲得更高的穩定性和更快的學習速度。

The benefits of using the advanced weightadjustment formulas include higher stability and faster speeds in the trainingprocess of the neural network.

穩定和速度對於深度學習尤其困難，因為它很難訓練。

These characteristics are especiallyfavorable for Deep Learning as it is hard to train.

本節只討論包含動量的訓練方法，這種方法已經使用很長時間了。

This section only covers the formulas thatcontain momentum, which have been used for a long time.

If necessary, you may want to study thisfurther with the link shown in the footnote.

雖然這種影響隨著時間的推移而減小，但舊的權重更新仍然一直存在。

Although the influence diminishes overtime, the old weight updates remain in the momentum.

因此，權重不完全受某些特定權重更新值的影響。

Therefore, the weight is not solelyaffected by a particular weight update value.

這樣可以使得網路的學習穩定性提高了。

Therefore, the learning stability improves.

此外，隨著權重更新的增加，動量也越來越大。

In addition, the momentum grows more andmore with weight updates.

因此，權重更新也會越來越大。

As a result, the weight update becomesgreater and greater as well.

從而提高了網路的學習速度。

Therefore, the learning rate increases.

——本文譯自Phil Kim所著的《Matlab Deep Learning》

更多精彩文章請關注微訊號：