There are two major types of uncertainty one can model. Aleatoric uncertainty captures noise inherent in the observations. On the other hand, epistemic uncertainty accounts for uncertainty in the model -- uncertainty which can be explained away given enough data. Traditionally it has been difficult to model epistemic uncertainty in computer vision, but with new Bayesian deep learning tools this is now possible. We study the benefits of modeling epistemic vs. aleatoric uncertainty in Bayesian deep learning models for vision tasks. For this we present a Bayesian deep learning framework combining input-dependent aleatoric uncertainty together with epistemic uncertainty. We study models under the framework with per-pixel semantic segmentation and depth regression tasks. Further, our explicit uncertainty formulation leads to new loss functions for these tasks, which can be interpreted as learned attenuation. This makes the loss more robust to noisy data, also giving new state-of-the-art results on segmentation and depth regression benchmarks.
核心贡献 · Key contributions
提出一个统一的贝叶斯深度学习框架，结合了输入依赖的偶然不确定性和认知不确定性。 Proposes a unified Bayesian deep learning framework combining input-dependent aleatoric and epistemic uncertainty.
推导出回归和分类的新损失函数，作为学习衰减，提高对噪声数据的鲁棒性。 Derives new loss functions for regression and classification that act as learned attenuation, improving robustness to noisy data.
在语义分割（CamVid、NYUv2）和深度回归（Make3D、NYUv2 Depth）基准上取得最先进结果。 Achieves state-of-the-art results on semantic segmentation (CamVid, NYUv2) and depth regression (Make3D, NYUv2 Depth) benchmarks.
表明偶然不确定性捕捉固有噪声并在大数据场景下提升性能，而认知不确定性检测分布外样本。 Shows aleatoric uncertainty captures inherent noise and improves performance in large-data regimes, while epistemic uncertainty detects out-of-distribution examples.
证明结合两种不确定性优于单独建模，且偶然不确定性几乎不增加计算量。 Demonstrates that combining both uncertainties outperforms modeling either alone, with minimal added compute for aleatoric uncertainty.
局限 · Limitations
通过蒙特卡洛丢弃法建模认知不确定性计算成本高，50 个样本导致高达 50 倍减速。 Epistemic uncertainty via Monte Carlo dropout is computationally expensive, causing up to 50x slowdown for 50 samples.
仅靠偶然不确定性无法检测分布外样本，限制了其在安全关键应用中的使用。 Aleatoric uncertainty alone cannot detect out-of-distribution examples, limiting its use in safety-critical applications.
该框架假设高斯或拉普拉斯似然，可能不适用于所有视觉任务。 The framework assumes Gaussian or Laplacian likelihoods, which may not be optimal for all vision tasks.
评估仅限于语义分割和深度回归，未探索对其他视觉任务（如目标检测）的适用性。 Evaluation is limited to semantic segmentation and depth regression; applicability to other vision tasks (e.g., object detection) is not explored.
该方法需要仔细调整丢弃概率和蒙特卡洛样本数量以达到最佳性能。 The method requires careful tuning of dropout probability and Monte Carlo sample count for optimal performance.
论文章节 · Sections（共 17）
摘要Abstract
1 引言1 Introduction
2 相关工作2 Related Work
2.1 贝叶斯深度学习中的认知不确定性2.1 Epistemic Uncertainty in Bayesian Deep Learning