Maximum Likelihood Uncertainty Estimation: Robustness to Outliers
This work addresses robustness in uncertainty estimation for regression tasks, which is important for reliable AI applications, but it is incremental as it builds on existing maximum likelihood methods with a specific distribution change.
The paper tackled the problem of degraded performance and incorrect uncertainty estimation in regression tasks due to outliers in training data, and found that using a heavy-tailed Laplace distribution improved robustness, providing better uncertainty estimates, separation for out-of-distribution data, and adversarial attack detection.
We benchmark the robustness of maximum likelihood based uncertainty estimation methods to outliers in training data for regression tasks. Outliers or noisy labels in training data results in degraded performances as well as incorrect estimation of uncertainty. We propose the use of a heavy-tailed distribution (Laplace distribution) to improve the robustness to outliers. This property is evaluated using standard regression benchmarks and on a high-dimensional regression task of monocular depth estimation, both containing outliers. In particular, heavy-tailed distribution based maximum likelihood provides better uncertainty estimates, better separation in uncertainty for out-of-distribution data, as well as better detection of adversarial attacks in the presence of outliers.