Statistical Approaches to Understanding Modern ML Methods
Aug 24, 2021
University of Wisconsin–Madison
When we use modern machine learning (ML) systems, the output often consists of a trained model with good performance on a test dataset. This satisfies some of our goals in performing data analysis, but leaves many unaddressed — for instance, we may want to build an understanding of the underlying phenomena, to provide uncertainty quantification about our conclusions, or to enforce constraints of safety, fairness, robustness, or privacy. As an example, classical statistical methods for quantifying a model’s variance rely on strong assumptions about the model — assumptions that can be difficult or impossible to verify for complex modern ML systems such as neural networks.
This workshop will focus on using statistical methods to understand, characterize, and design ML models — for instance, methods that probe “blackbox” ML models (with few to no assumptions) to assess their statistical properties, or tools for developing likelihoodfree and simulationbased inference. Central themes of the workshop may include:


Using the output of a ML system to perform statistical inference, compute prediction intervals, or quantify measures of uncertainty

Using ML systems to test for conditional independence

Extracting interpretable information such as feature importance or causal relationships

Integrating likelihoodfree inference with ML

Developing mechanisms for enforcing privacy, robustness, or stability constraints on the output of ML systems
 Exploring connections to transfer learning and domain adaptation
 Automated tuning of hyperparameters in blackbox models and derivativefree optimization

More information at https://ifds.info/ifdsmadlabworkshop/.