Dissertation explores bad minima in transfer learning via adversarial attacks
Original title: Exploring Bad Minima in Transfer Learning
Adversarial attacks on transfer learning reveal a strong link between source generalization and successful transfer.
Transfer learning applies knowledge from a source task to a related target task in deep learning. The dissertation hypothesizes that good generalization on the source task does not necessarily imply transferability and tests this via three types of adversarial attacks. Empirical results show a strong relationship between the two, though generalization remains potentially sufficient for transfer.
- Hypothesis: source generalization does not guarantee transferability to related tasks.
- Three types of adversarial attacks tested on transfer procedure.
- Results demonstrate strong correlation but do not rule out sufficiency of generalization.
From the original article
1 min read at arxiv.orgTransfer learning is a highly successful learning paradigm in deep learning which leverages knowledge from one task and applies it to another. Despite its success, the necessary conditions for knowledge transfer are unclear and require further research.
By Mina Remeli · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.