Yael Konforti

About

I am a PhD student at the University of Cambridge, supervised by Prof. Mateja Jamnik and Dr. Mateo Espinosa Zarlenga. I am affiliated with St Edmund’s College and the Department of Computer Science and Technology. I study how interpretability methods, especially concept-based ones, can expose the vulnerabilities of AI models and make them more reliable and safe to deploy.

Before my PhD, I spent several years in the industry as a Research Scientist at SightX LTD, building real-time object detectors and their evaluation protocols, and as an ML Researcher at BioMedSU, working on drug discovery from RNA-seq data.

Research

I am interested in the internal representations of AI models, where models fail, and how interpretability can help us understand why and fix it:

• Discovering and explaining failure modes: identifying subpopulations on which the model underperforms, using human-understandable concepts, and explaining the root causes of the failures [CB-SLICE].
• Removing spurious concepts: using sparse autoencoders to identify and suppress spurious embeddings of trained models (ongoing).
• Understanding model internals: modelling the inference path that an image, or a whole class, takes through the network as a probabilistic decision chain of visual words [Inference Graphs] [SIGN].

🎺 I’m open to advising MPhil projects on interpretability and bias mitigation.
Cambridge students, please see this link for details.

News