Data-Efficient Visual Inductive Priors

Overview

A retrospective examination of deep learning approaches for training with limited data, analyzing results from the VIPriors workshop series. This research reveals successful strategies for data-efficient learning without transfer learning.

Key Contributions

• Analysis of VIPriors workshop series on data-efficient deep learning
• Identifies successful patterns: model ensembles mixing Transformers and CNNs
• Heavy data augmentation as key to success
• Training from scratch without transfer learning on limited data

Links

• arXiv
• PDF