MLPs with a lot of layers

• One of the (maybe the) big successes of AI in the 80s was the MLP .
• It's still widely used today in analytics.
• In these nets, people only typically use one hidden layer, and at most two.
• Why? It's the vanishing gradient problem.
• It's trained by the backprop rule (or some variant).
• Backprop can only correct from the backpropagation of error.
• It doesn't work as well in the connections further from the output (and thus the error).