Efficient training of interval Neural Networks for imprecise training data.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Quantifying uncertainty in Neural Networks (NNs) is crucial for reliable decision-making.
• Existing methods for uncertainty quantification can be computationally intensive or lack robustness.
• Handling epistemic uncertainty, especially in large datasets, remains a significant challenge.

Purpose of the Study:

• To develop a robust and computationally feasible method for training NNs with interval predictions.
• To quantify the uncertainty associated with NN predictions.
• To address epistemic uncertainty in large datasets.

Main Methods:

• A novel backpropagation algorithm for NNs with interval predictions.
• Minimizing the maximum batch error at each step for numerical stability.
• Utilizing the non-convex Scenario approach for chance-constrained optimization post-training.

Main Results:

• The proposed method effectively trains NNs for interval predictions and uncertainty quantification.
• The approach accommodates uncertainty in training data, including adversarial examples.
• Computational complexity scales favorably with O(M⋅Niter), independent of dataset size, unlike other Interval Predictor Model methods.

Conclusions:

• This work presents the first computationally feasible approach for convex set-based epistemic uncertainty in large datasets.
• The method provides reliable predictions and avoids complex penalty function methods.
• The proposed technique enhances the trustworthiness and applicability of NNs in critical domains.