Towards Robust Low-Resource Fine-Tuning with Multi-View Compressed Representations