Quantifying calibration error in modern neural networks through evidence based theory

Artificial Intelligence (AI) systems, particularly neural networks, are increasingly employed in critical applications such as healthcare, finance, and autonomous systems. These systems play an integral role in decision-making, where the trustworthiness of their predictions becomes paramount. Trustworthiness in AI encompasses attributes like reliability, robustness, fairness, and transparency, yet these qualities are often difficult to evaluate, particularly in neural networks, which are typically viewed as "black-box" models. This opacity raises significant concerns about their trustworthiness, especially in sensitive domains where incorrect decisions can lead to severe consequences. Traditional performance metrics like accuracy, precision, and recall measure only the correctness of the model's predictions but fail to capture the confidence and uncertainty associated with those predictions. Confidence calibration, which aligns predicted probabilities with actual outcomes, has emerged as an important tool to address these shortcomings. Well-calibrated models provide predictions where the predicted probability corresponds to the actual likelihood of the event, ensuring that a 70% confidence means the event occurs approximately 70% of the time. However, despite its utility, calibration alone does not fully address the issue of trustworthiness, as it does not account for subjective uncertainty or provide an interpretable way to assess trust across a range of predictions. To address this, we propose the use of subjective logic for trustworthiness quantification in neural networks.