Guanjie Cheng, Boyi Li, Lingyu Sun, Mengying Zhu, Yangyang Wu, Xinkui Zhao, Shuiguang Deng
TADS is a new framework for selecting high-quality, task-relevant data for multi-task multimodal pre-training, improving efficiency and performance using less data.
The research introduces TADS, a system designed to improve how large-scale AI models are trained with diverse data. When training models that understand both images and text, like CLIP, it's crucial to use high-quality data. However, data from the internet can be messy and inconsistent. TADS selects the best data by considering its quality, relevance to specific tasks, and diversity, using a smart learning approach. Tests show that TADS can achieve better results using less data, making it a more efficient way to train AI models.
