Model training is the process by which neural networks are fit to a set of training data.
During training:
- data is input into the network and an output is generated via forward propagation.
- a loss function is calculated using the predicted output and actual label.
Not in the SD
The gradient of the loss function is computed using backpropagation, which is then used to optimise the weights in order to minimise the loss via gradient descent. This involves derivatives, and Euler’s method is commonly involved.