Robustness in Machine Learning: Explainability, Uncertainty, and Intervenability

#machine-learning #uncertainty #faculty #students

Neural network–driven AI systems, including large-scale foundation models, suffer from critical vulnerabilities: they behave unpredictably under unseen conditions, confidently produce inaccurate outputs, and resist meaningful human oversight. Traditional robustness measures—simple invariance to noise and perturbations—are insufficient for real-world deployment. This talk presents a human-centric framework that decomposes robustness into three actionable pillars: explainability, uncertainty quantification, and intervenability. We argue that every prediction made by a neural network must be accompanied by contextual explanations, calibrated confidence estimates, and mechanisms for human intervention.

We introduce our explanatory paradigms framework, rooted in abductive reasoning, which enables networks to answer correlative, counterfactual, and contrastive questions through gradient-based traversal of representation spaces, without retraining. Building on this, our introspective learning approach lets networks reflect on their own decisions, yielding features that unify explainability with uncertainty quantification. We then present our experts-in-the-loop framework for intervenability, which engages diverse domain specialists at each stage of the AI deployment cycle, from data curation to evaluation, supported by research in active learning under distributional shift, patient-aware active learning, and cost-centered annotation strategies. Applications span autonomous vehicle perception, computational ophthalmology, and seismic interpretation.

• Add Event to Calendar
  iCal
  Google Calendar

Loading virtual attendance info...

• 1889 Museum Rd
• University of Florida
• Gainesville, Florida
• United States 32611
• Building: Malachowsky Hall
• Room Number: 1000

---

• Contact Event Host

• Professor Ivan Ruchkin <iruchkin@ece.ufl.edu>

• Co-sponsored by UF ECE Department

Seminar