The success of a learning algorithm depends on its ability to balance between fitting the data and avoiding overfitting.
The best architecture comes from constraints, not from freedom.