On Pretraining Data Diversity for Self-Supervised Learning

arXiv:2403.13808v1 Announce Type: cross
Abstract: We explore the impact of training with more diverse datasets, characterized by the number of unique samples, on the performance of self-supervised learning (SSL) under a fixed computational budget. Our findings consistently demonstrate that increasing pretraining data diversity enhances SSL performance, albeit only when the distribution distance to the downstream data is minimal. Notably, even with an exceptionally large pretraining data diversity achieved through methods like web crawling or diffusion-generated data, among other ways, the …

arxiv cs.ai cs.cv cs.lg data data diversity diversity pretraining self-supervised learning supervised learning type