Overfitting occurs when details are modeled that aren't justified by the data. i.e. not statistically significant.

Dropping nodes seems a roundabout way of addressing this, and you'd get a similar effect by training less.