Trustworthy Machine Learning and Artificial Intelligence Frameworks for Scientists

Some scientists hesitate to use artificial intelligence (AI) and machine learning (ML) methods due to the lack of reproducibility, explainability, and transparency in these models; these qualities are collectively known as "trustworthiness". Trustworthy AI frameworks can help overcome this hesitancy by evaluating AI models beyond performance on a test dataset. Trustworthy AI frameworks for fields such as computer vision, natural language processing, and health care may include social responsibility aspects. In addition to these aspects, a Trusted AI framework for scientific ML models should additionally seek the model's agreement with physical laws of nature. However, a disconnect can arise between emphasized aspects of Trustworthy AI and the resources available to an AI/ML practitioner who wants to verify trustworthiness. In this talk, an overview of several available Trustworthy AI frameworks is presented in a scientific context. This is supported by a demonstration of some simple, general approaches for quantifying model trustworthiness, and work towards a unifying trustworthy AI framework and toolkit for physical scientists is presented.