Hands-On Machine Learning for Algorithmic Trading

Dropout

Dropout refers to the randomized omission of individual units during forward or backward propagation with a given probability. As a result, the omitted units do not contribute to the training error or receive updates.

Dropout is a very popular regularization technique because it is computationally very inexpensive and does not significantly constrain the type of model or training procedure that can be used. A downside is that more iterations are necessary to achieve the same amount of learning, but on the plus side, each iteration is faster due to lower computational cost.

It works as a regularization technique because it prevents units from co-adapting to or compensating for mistakes made by other units during the learning process, ...