Learning RNN weights

In practice, the state of the network is much more complex than a vector containing a weight for each class, as in the previous example. The weights of W_input, W_rec, and V cannot be engineered by hand. Thankfully, they can be learned through backpropagation. This technique was detailed in Chapter 1, Computer Vision and Neural Networks. The general idea is to learn the weights by correcting them based on the errors that the network makes.