Divisive Decisions: Improving Salience-Based Training for Generalization in Binary Classification Tasks