The project provides tools for assisting design of statistical models and Bayesian machine learning models. One key step in designing these models is the choice of the prior distribution, which is often extremely challenging. Prior elicitation is a known technique for assisting the choice, but the current tools are severely limited and often not used in practice. We provide practical open tools that considerably extend the capabilities and applicability of prior elicitation, in form of a computationally feasible parameterisation of joint prior distributions and practical means for eliciting such priors from an expert via preferential interaction that only relates to observable quantities, not directly to model parameters. The solutions work for both classical statistical models as well as for Bayesian machine learning models, including deep neural networks, improving the development pace and quality of the models in research and practical use.

2024

2028