Thema der Dissertation:
Proxy Gradient Estimation for Black-Box Feature Attribution Thema der Disputation:
A Modality-Agnostic Feature Attribution Framework via Proxy Gradient Estimation
Proxy Gradient Estimation for Black-Box Feature Attribution Thema der Disputation:
A Modality-Agnostic Feature Attribution Framework via Proxy Gradient Estimation
Abstract: Feature attribution explains machine decisions by quantifying the contribution of each input component. A widely used line of research relies on gradient information to determine attribution scores and produces quality explanations, benefiting from the internal access to the target model. In this talk, I will show how the idea of gradient-based explanation can be generalized to a black-box setting through gradient estimation, largely improving the flexibility and applicability of gradient-based solutions while preserving explanation quality. Specifically, the proposed feature attribution framework based on a proxy gradient estimator will be introduced. It bridges gradient-based and game-theoretic approaches, representing two popular categories of explainability research, and demonstrates great potential in the emerging multimodal studies.
Time & Location
May 11, 2026 | 02:00 PM
Seminarraum 006
(Fachbereich Mathematik und Informatik, Takustr. 9, 14195 Berlin)
&
WebEx