English

On consequences of finetuning on data with highly discriminative features

Machine Learning 2023-11-17 v2

Abstract

In the era of transfer learning, training neural networks from scratch is becoming obsolete. Transfer learning leverages prior knowledge for new tasks, conserving computational resources. While its advantages are well-documented, we uncover a notable drawback: networks tend to prioritize basic data patterns, forsaking valuable pre-learned features. We term this behavior "feature erosion" and analyze its impact on network performance and internal representations.

Keywords

Cite

@article{arxiv.2310.19537,
  title  = {On consequences of finetuning on data with highly discriminative features},
  author = {Wojciech Masarczyk and Tomasz Trzciński and Mateusz Ostaszewski},
  journal= {arXiv preprint arXiv:2310.19537},
  year   = {2023}
}

Comments

NeurIPS 2023 -- UniReps Workshop