PhD Student, Research Assistant in Explainable AI
Freie Universität Berlin · Berlin, Germany
2022 – Present
I hold a doctorate (Dr. rer. nat.) in Computer Science from Freie Universität Berlin, where my research centered on responsible AI, with a particular emphasis on the explainability and reliability of machine learning and general AI systems. Before that, I worked on related topics at the L3S Research Center at Leibniz Universität Hannover.
Building on several published works in feature attribution, my long-term research vision is to advance explainability methodologies that move beyond individual feature importance and provide deeper insights into the behavior of increasingly complex AI systems, e.g., LLMs and VLMs. In particular, I am now exploring higher-order feature interactions, data attribution, and the use of explanatory information as practical tools for improving the reliability, transparency, and sustainability of deep learning systems.
Experience
Academic roles
Research Assistant in Explainable AI
Leibniz Universität Hannover · Hannover, Germany
2020 – 2022
Student Research Assistant in Semi-supervised Machine Learning
Leibniz Universität Hannover · Hannover, Germany
2017 – 2018
Research & Publications
Selected work
Rethinking Explanation Evaluation under the Retraining Scheme
Revisits widely-used retraining-based evaluation for feature attribution, demonstrates its pitfalls, and proposes alternative efficient, distortion-free schemes for assessing explanation quality.
GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation
Derives a feature attribution framework based on proxy gradient estimation. The proposed method provides an unbiased estimation of Shapley values, which is generally applicable to different model architectures and data modalities.
On Gradient-like Explanation under a Black-box Setting
Explores the possibility of delivering gradient-like attributions with only query-level access. The resulting preserves a set of promising axioms of the white-box approaches and demonstrates competitive performance in the empirical experiments
For a full list of publications, please visit my
Google Scholar profile.
Projects
The projects that fund(ed) my research
Opportunities and risks of generative AI in cybersecurity (AIgenCY)
AIgenCY shall explore fundamental research on existing and forthcoming threats for and through generative AI, particularly large language models and foundation models. At the same time, suitable measures are to be developed to improve the detection of and defence against such cyber attacks.
Center for Trustworthy AI (ZVKI)
The independent think tank iRights.Lab founded the centre in partnership with the institutes Fraunhofer AISEC and IAIS as well as the Freie Universität Berlin. As a forum for debate in Germany, the CTAI makes the developments surrounding societal questions about artificial intelligence and algorithmic systems tangible.
Responsible Artificial Intelligence
Artificial intelligence (AI) technologies are the driving force behind digitization. Due to their enormous social relevance, a responsible use of AI is of particular importance. The research and application of responsible AI is a very young discipline and requires the bundling of research activities from different disciplines in order to design and apply AI systems in a reliable, transparent, secure and legally acceptable way.
KISWIND
The demand for automated damage detection in the field of civil infrastructures is high, for economic and safety reasons. This collaborative research project aims to contribute to a further development of automated damage detection in wind energy turbines based on acoustic emission testing (AET) and machine learning.
Contact
Reach out for collaborations, talks, or questions
Email:
yi.cai@fu-berlin.de
GitHub:
github.com/caiy0220
Scholar:
Google Scholar profile
LinkedIn:
LinkedIn profile
Visitors
Recent page visits by region