On Pretraining Data Diversity for Self-Supervised Learning

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Hammoud, Hasan Abed Al Kader, Das, Tuhin, Pizzati, Fabio, Torr, Philip, Bibi, Adel, Ghanem, Bernard
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866907983241281536
author Hammoud, Hasan Abed Al Kader
Das, Tuhin
Pizzati, Fabio
Torr, Philip
Bibi, Adel
Ghanem, Bernard
author_facet Hammoud, Hasan Abed Al Kader
Das, Tuhin
Pizzati, Fabio
Torr, Philip
Bibi, Adel
Ghanem, Bernard
contents 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 distribution shift remains a challenge. Our experiments are comprehensive with seven SSL methods using large-scale datasets such as ImageNet and YFCC100M amounting to over 200 GPU days. Code and trained models are available at https://github.com/hammoudhasan/DiversitySSL
format Preprint
id arxiv_https___arxiv_org_abs_2403_13808
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle On Pretraining Data Diversity for Self-Supervised Learning
Hammoud, Hasan Abed Al Kader
Das, Tuhin
Pizzati, Fabio
Torr, Philip
Bibi, Adel
Ghanem, Bernard
Computer Vision and Pattern Recognition
Artificial Intelligence
Machine Learning
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 distribution shift remains a challenge. Our experiments are comprehensive with seven SSL methods using large-scale datasets such as ImageNet and YFCC100M amounting to over 200 GPU days. Code and trained models are available at https://github.com/hammoudhasan/DiversitySSL
title On Pretraining Data Diversity for Self-Supervised Learning
topic Computer Vision and Pattern Recognition
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2403.13808