Interpretable and Robust Artificial Intelligence
Abstract
We proposed safe AI systems. These AI systems are trusted because they are interpretable and resilient because they are robust. In particular, we pushed the state of the art in modern probabilistic modeling, including probabilistic models for causal inference. Probabilistic models provide a natural way for domain experts to express their assumptions and then to derive algorithms to compute under those assumptions. The results are interpretable because, for each model, we have a clear mathematical understanding of what is assumed, how the structure of the data interacts with the inferences, and the boundaries of what can and cannot be captured. With the methods we developed around causality, the resulting system will also be robust---robust to changes in the world and to interventions. To directly aid scientific discovery, we studied our methods on several problems in medical informatics, cancer therapy analysis, computational biology, and statistical astrophysics.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 2023
- Accession Number
- AD1193162
Entities
People
- David M. Blei
Organizations
- Columbia University