Align Representations with Base: A New Approach to Self-Supervised Learning