Learning from Aggregate responses: Instance Level versus Bag Level Loss Functions