VLMixer: Unpaired Vision-Language Pre-training via Cross-Modal CutMix (ICML 2022) - View it on GitHub