In this white paper, artificial intelligence refers to a human-made system that is capable of intelligent behaviour in the sense that it develops the rules of its behaviour itself, based on a given environment (input data) and a given goal. Machine learning (see below) can be one part of such an AI system.
The black box metaphor is usually used for systems where the internal mechanisms are unknown, and only input and output can be observed. An ML model is thus referred to as a ‘black box’ model when it is not possible (or too difficult to be feasible) to explain to a human how a given input data leads to the output data provided by the model (prediction). Common examples include neural networks and ensemble models. Note that there are different interpretations of ‘explainable’ in this context: In this paper, a theoretical possibility to calculate the model output manually from a large number of internal model parameters is not deemed to be a sufficient ‘explanation’ if that procedure does not lead to a set of rules that is comprehensible by humans. Approaches to make the relation between model input and output more comprehensible by adding an additional, separate algorithm (such as lime, shapley values) do not turn a black box model into a white box model.
An independent review of an organisation’s adherence to regulatory guidelines. (See  for details.)
Hyperparameters are used to control the learning process in machine learning. These parameters are controlled by the coder, unlike the values of the model parameters which are derived via training.
Machine learning is a field of computer science dealing with methods to develop (‘learn’) rules from input data to achieve a given goal. A machine learning algorithm is a certain programmatic implementation of a strategy to find such rules (for example, with a neural network, a logistic regression or a decision forest). A machine learning model is the resulting set of rules (encoded in model parameters), inherently including the type of the ML algorithm. The ML model can be used to make predictions on data previously unknown to the ML model.
In the case of an SAI, a performance audit is an independent evaluation of a policy, programme or institute of a country’s central government. The aim of this specific type of audit is to assess the efficiency and effectiveness of the spend resources. In general there is an adherence to previously agreed methodology and an objective and systematic execution of the evaluation.
In the context of ML, the ‘white box’ metaphor is used to emphasise that the model outcome can be explained to humans given the input data and the model parameters. Common examples are logistic regression and (single) decision trees.
Initially training data is used to fit an optimal performing ML model. The validation data is then used to provide an evaluation of the model fit whilst tuning the hyperparameters. The test data is used last in order to provide a final evaluation of the model’s performance.