Reproduce “Label Noise Robust Image Representation Learning based on Supervised Variational Autoencoders in Remote Sensing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gencer Sumbul, Begüm Demir
A label noise robust image representation learning method based on Supervised Variational Autoencoders (SVAE) is introduced for remote sensing, demonstrating consistently higher F1-scores and Overall Accuracy across multi-label classification and semantic segmentation tasks with up to 60% synthetic label noise.
Key insights
- Supervised Variational Autoencoders are inherently less susceptible to annotation noise compared to non-generative deep neural networks, providing a robust baseline for comparison.
- The discrepancy between the task losses of a primary DNN and an SVAE on a given sample serves as an effective signal for detecting and quantifying the likelihood of label noise.
- A dynamic importance reweighting scheme, where noisy samples are down-weighted more heavily in early training stages, effectively mitigates overfitting to corrupted labels and improves learned representation quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.