Rather than training the entire network at the same time, this technique involves training one layer at a time using a greedy approach. This is where the broader problem is broken down into smaller subproblems, and then the optimal solution to each subproblem is combined to make the solution to the larger problem. Sometimes this takes the form of training the input layers first and then moving through the network, and sometimes the layers near the input and output are trained and new hidden layers are added in the middle of the two.