Measure Twice, Cut Once: Quantifying Bias and Fairness in Deep Neural Networks