Hands-On Artificial Intelligence for Beginners

Backpropagation through time

RNN utilizes a special variation of regular backpropagation called backpropagation through time. Like regular old backpropagation, this process is often handled for us in TensorFlow; however, it's important to note how it differs from standard backpropagation for feedforward networks.

Let's recall that RNNs utilize small copies of the same network, each with its own weights and bias factors. When we backpropagate through RNNs, we calculate the gradients at each time step, and sum the gradients across the entire network when computing the loss. We'll have a separate gradient from the weight that flows to the computation that happens at each of the time steps, and the final gradient for W will be the sum of the ...