Lightweight parameter-efficient adaptation and discriminative sample-aware loss function improve few-shot learning performance.