Resolve high training loss in overfitted MLPs by fixing data preprocessing, architecture capacity, and learning rate settings.