A naukri.com initiative
Arxiv
1d
256
Image Credit: Arxiv
On the Out-of-Distribution Generalization of Self-Supervised Learning
- This paper focuses on the out-of-distribution (OOD) generalization of self-supervised learning (SSL).
- Analysis of mini-batch construction in SSL training reveals one explanation for SSL's OOD generalization.
- SSL learns spurious correlations during training, leading to a decrease in OOD generalization.
- To address this issue, a post-intervention distribution (PID) grounded in the Structural Causal Model is proposed, along with a batch sampling strategy enforcing PID constraints for optimal worst-case OOD performance.
Read Full Article
15 Likes
For uninterrupted reading, download the app