<ul><li>Label Distribution Learning (LDL) has promising representation capabilities for characterizing polysemy but suffers from complexity and high cost of label distribution annotation.</li><li>To address uncertainty arising from inexact labels in LDL, a Latent Label Distribution Grid (LLDG) is proposed to create a low-noise representation space.</li><li>LLDG models uncertainty by constructing a label correlation matrix and expanding values into Gaussian distribution vectors.</li><li>LLDG-Mixer is utilized to reconstruct LLDG, enforcing a customized low-rank scheme to reduce noise in label relations, demonstrating competitive performance in classification tasks.</li></ul>

Latent label distribution grid representation for modeling uncertainty

Discover more