One of CASMI’s goals is to develop repeatable and operational processes for the identification and mitigation of negative impacts of machine learning applications and the causes of those impacts. To this end, we have developed an evolving evaluation framework to guide the work and vision.
The evaluation framework provides a foundational structure for the design and evaluation of machine learning applications by decoupling fact-finding from evaluation. The framework divides the task of evaluating the human impacts of machine learning (ML) systems into two phases:
- Fact-finding related to three primary components:
- Data and how it was sourced and manipulated
- Central ML algorithms and how they were applied
- How the resulting systems are designed to interact with human users
- The evaluation phase examines how, given those facts, the system impacts the goals and values associated with a particular domain or field of use.
The framework is also the starting point for CASMI’s research roadmap, a set of specific research problems necessary to further operationalize the design, development, and evaluation of AI systems from the perspective of human health and safety.