EVENT DOSSIER
To address the underfitting problem at the tail end due to ignoring the heteroscedasticity of long-tailed data in deep unbalanced regression (DIR) tasks, researchers proposed the Decoupled Uncertainty Optimization (DUO) framework. This framework models instance-level uncertainty using conditional Gaussian distributions, converts it into dynamic enhancement signals for tail samples through decoupled mean-variance optimization, and introduces a distribution-overlap-based contrastive learning mechanism to mitigate feature looseness and semantic entanglement. In DIR benchmark tests in visual and biological fields such as IMDB-WIKI-DIR, AgeDB-DIR, and AAV2-DIR, the DUO framework achieved the best few-shot bMAE and GM metrics, while also maintaining competitiveness in few-shot MAE.
Deep Imbalanced Regression (DIR) 在年龄估计、深度预测等任务中广泛存在，现有方法常忽略长尾数据的异方差性导致尾部欠拟合。为此，研究者提出 DUO 框架，通过条件高斯分布建模实例级不确定性，并利用解耦均值 - 方差优化将其转化为尾部样本的动态增强信号；同时设计基于分布重叠的对比学习机制以缓解特征松散与语义纠缠。在 IMDB-WIKI-DIR、AgeDB-DIR 和 AAV2-DIR 等视觉及生物 DIR 基准测试中，DUO 取得了最佳的 few-shot bMAE 和 GM 指标，同时在 few-shot MAE 上保持竞争力。