Reproduce “DLME: Deep Local-flatness Manifold Embedding” on OpenResearch

About this paper

Zelin Zang, Siyuan Li, Di Wu, Ge Wang, Lei Shang, Baigui Sun, Hao Li, Stan Z. Li

DLME (Deep Local-flatness Manifold Embedding), developed by researchers from Zhejiang and Westlake Universities and Alibaba Group, reconstructs high-quality, locally flat low-dimensional embeddings by overcoming structural distortion and local collapse in high-dimensional data. The system achieves state-of-the-art performance across classification, clustering, and visualization tasks on diverse datasets, improving accuracy by over 5% on several metrics compared to other methods.

Key insights

• The Local Flatness Assumption, enforced by minimizing manifold curvature, provides a tighter geometric constraint than first-order similarity, yielding more discriminative and linearly separable embeddings.
• Leveraging different degrees of freedom in t-distribution kernels for structure and embedding spaces in the loss function is theoretically crucial for explicitly minimizing average local curvature.
• The DLME loss function acts as a smoother version of contrastive learning objectives, preventing local collapse and making contrastive learning benefits accessible to domains with limited data.

Summary by alphaXiv.