Dual Accuracy-Quality-Driven Neural Network for Prediction Interval Generation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Accurate uncertainty quantification is crucial for reliable deep learning (DL) model deployment.
• Prediction intervals (PIs) are essential for regression tasks, providing a range for predictions.
• High-quality PIs must be narrow and accurately capture probability density.

Purpose of the Study:

• To develop a method for automatically learning prediction intervals (PIs) for regression-based neural networks (NNs).
• To enhance the quality of PIs by ensuring they are narrow and cover the true values effectively.
• To improve the reliability of DL models in real-world regression applications.

Main Methods:

• Training two companion neural networks: one for target estimation and another for PI bounds.
• Designing a novel loss function for the PI network with objectives for minimizing PI width and ensuring PI coverage.
• Implementing a self-adaptive coefficient to balance the loss function's optimization objectives.

Main Results:

• The proposed method successfully generates prediction intervals with nominal probability coverage.
• The method produces significantly narrower PIs compared to state-of-the-art approaches.
• Target estimation accuracy is maintained without detriment, leading to higher-quality PIs.

Conclusions:

• The novel method effectively learns high-quality prediction intervals for deep learning regression models.
• The approach enhances model reliability by providing accurate and narrow uncertainty estimates.
• This work contributes to more trustworthy AI applications through improved uncertainty quantification.