Problems with Backprop

• Local minima

• can easily get stuck in a locally minimal error

• Overfitting

• can overlearn the training data
• lose the ability to generalise

• Sensitive to the MLP topology

• number of connections
• “free parameters”