Research Roadmap
The CASMI Research Roadmap is a set of research goals and an outline for the work necessary for operationalizing safety in intelligent systems. View the full Research Roadmap and the accompanying Evaluation Framework or read the Executive Summary below.
Executive Summary
Machine Learning (ML) and Artificial Intelligence (AI) are transformational technologies fueling gains in everything from processing data to fighting fires. Medical diagnostic systems, autonomous vehicles, and personal assistants have all incorporated the power of AI, but its rapid adoption increasingly outpaces our ability to understand its human impact. While many groups are focused on the fundamental articulation of ethical principles for AI, there is an urgent need for mechanisms to operationalize the principles of ethics in system development.
Northwestern University and UL Research Institutes established the Center for Advancing Safety of Machine Intelligence (CASMI) in response to this need. CASMI is a research hub founded to create best practices for safely designing, developing, and deploying Artificial Intelligence technologies. To further its mission of operationalized safety, CASMI is creating tools for evaluation and is identifying the necessary further research.
This Research Roadmap encompasses approximately 200 research questions that are necessary to guide the center’s current work and potential projects. Aimed at addressing areas where current knowledge and means are insufficient for evaluating system safety, these research questions were identified through the exercise of CASMI’s Evaluation Framework.
The Evaluation Framework outlines a mechanism to guide the assessment of system safety pre- and post-deployment. The Framework employs a decomposition approach, establishing the facts and issues of a system at the component level to then inform the system evaluation.
The Framework first progresses through a fact-gathering phase addressing the major areas in which system issues and research opportunities emerge:
- Data – Biases and other issues can arise throughout the data acquisition, cleaning, and normalization processes.
- Algorithm – The cycle of training and testing, the choice of specific algorithm, and the training and retraining dynamics each impact system performance and the nature of its application issues.
- Interaction – ML systems inherently interact with human users and the nature of that human interaction can determine much of the system’s impact.
The Framework then progresses to evaluation informed by questions in the humanities and social sciences.
- Evaluation – Identifying and measuring the human impacts of an ML application must be done relative to the goals and values of the domain in which it is deployed.
The intent of CASMI’s funded research at Northwestern and partner institutions is to further refine the roadmap or to address a key need in one of the Framework component areas. While the Evaluation Framework was a mechanism for developing the Roadmap, the Roadmap is a living document and an independent set of evolving research goals. As CASMI continues to uncover areas of opportunity, the Roadmap will be revised to prioritize research that advances the mission of operationalizing the safety of machine intelligence.
Back to top