Reproduce “Local Manifold Augmentation for Multiview Semantic Consistency” 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
Yu Yang, Wing Yin Cheung, Chang Liu, Xiangyang Ji
Local Manifold Augmentation (LMA) presents a data-driven augmentation method for self-supervised learning, utilizing an Instance-Conditioned GAN to generate diverse yet semantically consistent views. This approach improves linear classification accuracy by up to 11% and enhances robustness to various distribution shifts over traditional handcrafted augmentations.
Key insights
- Modeling local data distributions with an IC-GAN can effectively generate rich, semantically consistent augmentations that go beyond the capabilities of handcrafted rules.
- Probabilistic application of generative augmentations is critical to mitigate issues like mode collapse and integrate generated data synergistically with real data for performance gains.
- The quality of the underlying generative model significantly influences the effectiveness of LMA, indicating a direct correlation between GAN advancements and SSL performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.