Patrick Hemmer

Background

The rapid advances in Artificial Intelligence (AI) have enabled machine learning models to perform predictions as accurately as human experts in an increasing number of tasks. These developments fueled the vision of AI taking on an advisory role by providing human decision-makers with predictions and additional information such as explanations. However, in many application domains, e.g., fraud detection or credit scoring, the number of cases on which a decision has to be made is too large for sole human decision-making. As neither human experts nor AI models are free from false decisions, the paradigm of “learning to defer” as a particular form of human-AI collaboration has emerged. The idea is to understand humans and AI as a team and to train the AI to determine whether the human or AI should take over the final decision to achieve a superior team performance compared to either deciding on its own. However, current “learning to defer” frameworks entail several unfeasible requirements hindering their real-world applicability. Examples encompass the availability of human predictions in addition to ground truth labels, workload management to avoid too few or many instances being deferred or considering dynamic environments (Leitão et al., 2022).

Research Goal

For this reason, the thesis aims at studying how "learning to defer" frameworks can address additional requirements such as the ones mentioned above. As a first step, its objective is to implement state-of-the-art "learning to defer" frameworks and evaluate them on several scientific datasets. As a second step, the goal is to identify approaches from the literature suitable for extending existing frameworks, followed by a conceptualization and implementation to address one particular selected downside of existing "learning to defer" frameworks. Finally, we aim to evaluate the conducted extension experimentally to assess its utility.

In detail, you will…   

- gain theoretical knowledge and apply it to a practical use case.

- become an expert in human-AI collaboration.

- develop, implement, and evaluate human-AI systems that learn to allocate instances to human experts.

We look forward to receiving your application because you...

- have a solid technical understanding of machine and deep learning techniques.

- are proficient in Python programming (e.g., pandas, tensorflow, pytorch, scikit-learn).

- work in an independent and accurate way.

Details

Start: immediately | Duration: 6 Month

We offer you a challenging research topic, close supervision, and the opportunity to develop practical and theoretical skills. If you are interested, please send your CV, transcript of records, and a brief letter of motivation to patrick.hemmer@kit.edu.

References

Leitão, D., Saleiro, P., Figueiredo, M. A., & Bizarro, P. (2022). Human-AI Collaboration in Decision-Making: Beyond Learning to Defer. arXiv preprint arXiv:2206.13202.