Simultaneous Dimensionality Reduction: A Data Efficient Approach for Multimodal Representations Learning