Bias Mitigation Framework for Intersectional Subgroups in Neural Networks