Backpropagation

A method to tune the parameters a neural network. Often done by gradient descent, where the change in the loss function is calculated chain rule using computation graphs

$\frac{\partial J}{\partial {\Theta }^{(L)}}=(a^{(L)}-y)(a^{(L-1)}{)}^T$

• The derivative of the loss function with respect to the last layer’s parameters