Uncertainty Quantification
In the context of Machine Learning (ML), uncertainty refers to the lack of confidence in the results produced by a model. Looking at the notion of generalization beyond the data seen during the learning step, it is important to recognize that any decision made by an ML model is not provably certain but rather hypothetical. This notion stems from the process of induction, where the model infers patterns and rules based on observed data. Consequently, uncertainty becomes a crucial aspect to consider in ML models, contrasting with confidence. Several factors can influence the decisions made by a model and create uncertainty, such as insufficient or biased data, noise in the data, model architecture, and lack of interpretability, to name a few. Taking these factors into account and mitigating them improves uncertainty and increases the trustworthiness of model decisions. In statistical terms, uncertainty is encapsulated through standard probabilities and probabilistic predictions. The model’s ability to generalize beyond the data that it has learned from is a fundamental requirement, a process facilitated by extracting general principles from specific instances and subsequently applying them to unseen data. As a consequence, the predictions made by such a model are inherently uncertain. For this reason, the uncertainty in mathematical models stems from the stochastic perspective of the model’s output, in contrast to a deterministic output.