Algorithm Specific Methods

• Backprop - train several networks simultaneously and use a correlation penalty in the error function
• genetic operators to generate new network topologies - multiplicative term that incorporates the diversity of the classifiers - prune to N best networks
• training on auxiliary task as well as the main task - diverse classifiers can be learned with one primary task and but with different auxiliary tasks such as predicting one of its input features
• network whose secondary prediction is best is the winner - encourage different networks to become experts a predicting the auxiliary task in different local regions - causes the errors in the primary output to become decorrelated
• DT - option-trees - equivalent and more understandable than bagging