Okay, so now we have some good motivation for why this combination of nodes might help us make predictions. How are we actually going to adjust all of the sub-functions of our neural network nodes based on some input data? The answer is called backpropagation.
Backpropagation is a method for training our neural network that involves doing the following iteratively over a series of epochs (or exposure to our training dataset):
- Feeding our training data forward through the neural network to calculate an output
- Calculating errors in the output
- Using gradient descent (or other relevant method) to determine how we should change our weights and biases based on the errors
- Backpropagating these weight/bias changes into ...
