- 论文公开站arXiv
TCP_α:用于可靠音乐信息检索的边际控制置信度估计
$TCP_α$: Margin-Controlled Confidence estimation for reliable Music Information Retrieval
深度神经网络常过度自信,即使预测错误也给出高置信度。后验置信度估计通过训练轻量辅助头解决此问题,但现有目标存在歧义。本文提出TCP_α,一种新置信度目标,通过引入边际控制惩罚来分离正确与错误预测的置信度,并证明其分离边际与类别数无关且随惩罚参数单调增加。研究针对不平衡回归的训练策略,并在拉格识别等任务上验证有效性。
This paper proposes TCP_α, a novel confidence target for post-hoc confidence estimation that introduces a margin-controlled penalty to ensure complete separation between correct and incorrect predictions. It is evaluated on rāga identification and ornamentation detection, showing robustness under domain shift.
意义:为音乐信息检索等任务提供可靠的置信度估计,有助于开发者构建更可信的AI系统,减少因过度自信导致的错误决策。