Conducting a series of online and in-person workshops on ML safety and learning theory using probably approximately correct guarantees. The aim of these workshops is to introduce the listener to the notion of error-confidence bounds within ML and demonstrate practical applications of learning theory in deriving safety guarantees for learning-enabled systems.

Presentation Slides

The presentation slides used in these workshops are available here

PAC Guarantees

These workshops are divided into multiple parts: (1) guarantees set during the learning process and (2) post-training verification.

Guarantees Apriori

We take a more theoretical approach at placing bounds on the learning process based on the expected training distribution and the expected information yield of new samples with respect to training time. This allows us to bound the sample-complexity of training using error-confidence parameters.

Post-Training Verification

We take a more empirical approach at placing bounds on the outcome of training using a monte-carlo sampling based process and optimize using scenario based approaches to to find error given a confidence value.