Скачать или смотреть Less is More Data Specialization for Various Modalities

Speaker: Alla Barseghyan (YerevaNN)
Topic: Less is More? Data Specialization for various modalities
DataFest Yerevan 2025, https://datafest.am/

Abstract: What if your available compute is limited, while your dataset is far larger than you can process? What if the quality of your pretraining data is uncertain? And what if you actually need much less data than you have? Data specialization techniques can help identify a compact and effective subset of data, aligned with a given architecture, objective, and compute budget. In this work, we compare different data filtering algorithms across multiple modalities—including natural images (ImageNet), remote sensing imagery, and X-ray scans—to identify effective strategies for curating high-quality datasets.