On Patching Learning Discrepancies In Neural Network Training