QuickNets: Saving Training and Preventing Overconfidence in Early-Exit Neural Architectures