这篇博客探讨了预测模型中的不确定性，分为两种类型： aleatoric 和 epistemic 不确定性。aleatoric 不确定性源自数据内在的随机性，无法减少；而 epistemic 不确定性则因知识缺乏引起，可通过收集更多标签数据来降低。主动学习策略关注减少 epistemic 不确定性以提升预测模型性能。
1. Here, we have to distinguish between the aleatoric uncertainty that is caused by high Bayesian error, and the epistemic uncertainty, which is caused by a lack of information[22].
We are not able to reduce the aleatoric uncertainty, but we can acquire more labels to reduce the epistemic uncertainty in the currently considered neighborhood.
The prediction uncertainty can be decomposed into epistemic uncertainty and aleatoric uncertainty.
Epistemic uncertainty accounts for the uncertainty in a prediction model caused by a lack of knowledge. This uncertainty can be reduced by collecting more labeled instances. Conventional query selection in active learning is associated with the effective reduction of epistemic uncertainty to improve the prediction model[43].
Aleatoric uncertainty refers to the notion of randomness caused by the stochastic variability inherent in the data. Owing to its inherent nature, this uncertainty is irreducible.
[ ]Jongmin Han, Seokho Kang. Activelearning with missing values considering imputation uncertainty. KBS.2021
扫一扫
5117
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
