Problems with MLP

• Initialisation of weights

• Random initialisation can cause problems with training

• start in a bad spot

• Can initialise using statistics of training data

• PCA
• Nearest neighbour