Quantification of Uncertainties in Neural Networks

Artificial neural networks only compute point estimates and thus, do not provide the user with a proper decision space. In high-risk use cases, the confidence of the neural network is an important support for decision-making. Bayesian neural networks extend classical deep neural networks with a probability component and allow the user to assess the probability distribution over the prediction. Due to the large number of parameters to be learned, the calculation of the predictive probability can only be performed approximately in practice. In recent years, many methods have been developed to efficiently learn the parameter distributions for Bayesian neural networks. Each of these has different advantages and disadvantages, and thus can be used for different applications. Quantifying uncertainty in the context of neural networks allows the user to interpret the results more comprehensively as well as to assess the risk and therefore makes an important contribution to the user’s digital sovereignty.