Exploring Learned Representations of Neural Networks with Principal Component Analysis