Causality, responsibility, blame, and interventions, Hana Chockler

Hana Chockler (King’s College London)

Causality, responsibility, blame, and interventions – how can we use them in software engineering?

There is a number of challenges in software engineering that can be roughly classified as “explanations challenges”: why is this system correct? why is this a buggy trace of a program? I am getting an error – where is the fault? why did my neural network classify this image as a racoon? Not understanding why a certain decision has been made by an automated tool is not only worry-inducing for the users, but is also counter-productive for fixing errors. Indeed, if our automated tool says that the program contains an error and brings an erroneous execution as an example, we need more information in order to localise the problem and fix it. Similarly, if our neural network keeps misclassifying the inputs, we need to understand why this is happening in order to fix the problem. And even if our automated tool says that the program is correct, how can we be confident that the tool worked correctly? In this talk I will survey these questions, which serve as a motivation for the talk. I will then introduce the notions of actual causality (introduced by Halpern and Pearl), and its quantitative measures responsibility and blame, as well as the more recent notion of interventions in causal models. I will demonstrate that these notions can be useful to answer the questions above, and that the computation of causality can be integrated into software engineering tools to produce helpful information that increases the user’s confidence and aids in locating and fixing errors. I will conclude the talk with outlining some interesting future directions.