Continuous Learning

How can AI systems effectively continue learning from limited, potential biased, and evolving inspection data while maintaining focus on fairness and usefulness for inspectors and inspectees?

The deployment of AI systems to support inspectors creates a long-term relation between the AI system and the inspector community. In an optimal collaboration, these models continue learning beyond the initial training point based on a relatively low number of cases with known inspection results, often biased by experience-based selection process and/or practical circumstances. This may generate uncertain or erroneous predictions which can be mitigated. This work package will focus on three potentials for improved learning process of AI models.

Key Challanges in this domain

The first type of learning is to better exploit expert knowledge. An AI algorithm lacks general knowledge of the world and may lack complex visual or dynamic background of individual cases and can therefore make mistakes that would appear trivial to field experts. A process of expert augmented learning will help in reducing impacts of learning instances with incorrect or highly specialized characteristics.

Secondly, the system can actively identify new cases with a high expected contribution to learning in an explorative and exploitive way. As such, when there are a limited number of cases to evaluate, new cases can be identified, or a fitting evaluation strategy can be picked.

Thirdly, the use of the AI system, by successfully identifying high risk inspections, aims to drive behavioral change in the target population. The model should be able to adapt itself to this moving target and efficiently learn these changes in order to maintain its added value to the inspectors. During the PhD the practical aspect of learning curves for the lab will be researched to see if there are benefits for decision makers to adjust. The adjustment can be, but not limited to, the sampling method, classification method or a different data collection purpose.

Meet the researcher

Misja Groen

Utrecht University

I finished my bachelor hbo-ICT at Windesheim Zwolle and my Masters Business Informatics at Utrecht University. In my free time I like to play video games, listen to music and play some volleybal and dodgeball.