- Main
Contrastive, Causal, and Game Theoretic Explanations to Understand Multi-Modal Models
- Lahiri, Aditya
- Advisor(s): Salimi, Babal
Abstract
With the surge in use of machine learning and deep learning models for critical decision making in fields like healthcare, social justice and finance, the need to understand, debug and audit these blackbox systems has become pertinent. It has given rise to an entire field called Trustworthy and Responsible AI. Explainable AI is a subfield of this which focusses of explaining the decision making process of these highly complex blackbox systems. Our work concentrates on explanations that are causal and contrastive in nature. Instead of capturing spurious correlations, we aim to capture the underlying causal interactions in the model. We also want our explanations to be contrastive since as humans, we primarily understand outcomes and decision by comparing them to other similar situations. We rarely look at things in isolation, and rely on comparison and contrast to deepen our understanding. We incorporate notions from Game Theory to tackle the feature attribution problem of ascertaining how much difference an input feature makes on the outcome of an instance, and how that input feature value's difference in a given pair of instances affect their outcomes. Our methods address blackbox models that can be either trained on images or tabular data. They can help understand model decision making from the lens of multiple stakeholders- model builders, model auditors or model users. The comprehensive coverage of our proposed methods enable us to produce a toolkit for model explainability that is widely applicable and also generalizable.