Scaling Pre-training to One Hundred Billion Data for Vision Language Models
This work tackles the problem of building inclusive multimodal systems for diverse cultures and languages, which is significant for underrepresented communities and organizations seeking to develop culturally sensitive AI models.
The authors investigated pre-training vision-language models on 100 billion examples, finding that model performance saturates on common Western-centric benchmarks but achieves substantial gains on tasks of cultural diversity and low-resource languages, with improvements of up to 100 billion scale web data coverage of long-tail concepts. The results show that scaling to 100 billion examples is vital for building inclusive multimodal systems.
We provide an empirical investigation of the potential of pre-training vision-language models on an unprecedented scale: 100 billion examples. We find that model performance tends to saturate at this scale on many common Western-centric classification and retrieval benchmarks, such as COCO Captions. Nevertheless, tasks of cultural diversity achieve more substantial gains from the 100-billion scale web data, thanks to its coverage of long-tail concepts. Furthermore, we analyze the model's multilinguality and show gains in low-resource languages as well. In addition, we observe that reducing the size of the pretraining dataset via quality filters like using CLIP, typically used to enhance performance, may inadvertently reduce the cultural diversity represented even in large-scale datasets. Our results highlight that while traditional benchmarks may not benefit significantly from scaling noisy, raw web data to 100 billion examples, this data scale is vital for building truly inclusive multimodal systems.