In backwards propagation, one tries to minimize the cost function in the most efficient way by looking at small changes in the bias, the previous activation, and the weights of a neuron. We start in the output layer and work backwards towards the input layer. Is the cost function usually only defined using the output layer? Is it possible to define a cost function that uses a combination of the output layer and other layers (e.g. various hidden layers or even the input layer)?
In a sense, is "forward propagation" combined with "backwards propagation" possible?