Challenges and Open Research Questions
Modern Machine Learning has reached and continues to reach new, previously unknown highs in performance on an increasing number of tasks, often surpassing human accuracy in the process. Many of these tasks, such as object detection in images or natural language processing, are of high industrial relevance.
However, the deployment of these modern techniques is not straightforward as they sometimes suffer from immanent shortcomings like lack of robustness, lack of interpretability, or lack of theoretical guarantees. Moreover, some methods inherit and even tend to amplify flaws of the underlying training data, like bias and spurious correlations. This might lead to bias against people’s color of skin and gender or open doors to unwanted ways of extracting personal data. These shortcomings often impede the downstream analysis goals which the Machine Learning methods are intended to achieve namely, to draw robust conclusions from experimental data.
In the field of trustworthy Machine Learning, the scientific community investigates manifold ways to mitigate or to entirely avoid these drawbacks. The ML2R team plays an active role in this community.
Our approaches include:
- Adjusting model architectures and training
- Working towards a better theoretical understanding of data-driven ML
- Supporting end-users in the selection of ML methods via descriptions of properties including quality bounds, and resource requirements
- Measuring data quality and severity of the shortcomings
- Developing tools for human supervision or interaction with the aim of improving the learning process or achieving better explainability
- Investigating ways of combining connectionist and symbolic forms of ML
- Precisely imposing structure onto embedding spaces
Current Research Highlights
The ML2R team contributes to current lines of research in multiple ways. We put strong emphasis on the topic of model transparency, as ways to follow decisions are both dependent on the chosen model family and the application at hand. In particular, a group within the team is investigating the relationship between explainability and informed ML, looking into ways of integrating structured knowledge and taking advantage of this structure to create meaningful explanations.
Another group within the team aims at facilitating the use of Machine Learning for non-experts by deriving and summarizing key properties and guarantees for model families and training procedures. Eventually, this will provide a set of simple rules by which one can obtain safe and reliable ML models without highly experienced personnel and regardless of the actual application.
What is more, researchers from ML2R apply methods of interactive and informed ML to increase the robustness of data-driven models by including formal knowledge or human feedback. For this purpose, the researchers firstly rely on ex-post validation of model predictions by verifying scientific consistency or regulatory requirements. Secondly, they apply representation learning methods that target at embedding data along with prior knowledge.
Recommended Reading and Selected Publications
Introduction and Overview
G. Marcus, E. Davis: Rebooting AI: Building Artificial Intelligence We Can Trust. Pantheon Books, 2019.
Transparency, Explainability and Interpretability