Bridging the Gap: Unifying the Training and Evaluation of Neural Network Binary Classifiers

How can this training-evaluation gap be addressed? While specific techniques have been adopted to optimize certain confusion matrix based metrics, it is challenging or impossible in some cases to generalize the techniques to other metrics.