Simple and Effective Approaches for Uncertainty Prediction in Facial Action Unit Intensity Regression.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Estimating prediction uncertainty is crucial for reliable decision-making in critical applications.
• Existing uncertainty estimation methods in regression tasks vary in performance and complexity.

Purpose of the Study:

• To introduce and evaluate two novel, simple approaches for uncertainty estimation in regression.
• To compare the performance and complexity of these new methods against established techniques.

Main Methods:

• Operationalized uncertainty as the absolute error between prediction and ground truth.
• Developed two secondary models to predict the uncertainty of a primary predictive model.
• One approach used non-parametric methods based on similar observations; the other directly trained a secondary model for uncertainty prediction.

Main Results:

• Both proposed uncertainty estimation approaches outperformed established methods on MNIST, DISFA, and BP4D+ datasets.
• Directly predicting uncertainty showed superior performance compared to indirectly estimating it.
• The proposed methods offer a computationally efficient way to gauge prediction reliability.

Conclusions:

• The two novel methods provide effective and simple solutions for uncertainty estimation in regression.
• Direct uncertainty prediction is a more effective strategy than indirect estimation.
• These findings enhance the trustworthiness of machine learning models in real-world applications.