Causal Intersectionality and Dual Form of Gradient Descent for Multimodal Analysis: a Case Study on Hateful Memes

TL;DR

This paper explores how causal intersectionality and gradient-based methods can improve understanding and explanation of multimodal models, especially in detecting hateful memes, by framing the task as an Average Treatment Effect estimation.

Contribution

It introduces a novel approach combining causal intersectionality with gradient-based analysis to interpret multimodal models and demonstrates its application in hateful meme detection.

Findings

Hateful meme detection modeled as ATE estimation using intersectionality.

Gradient-based attention scores reveal model-specific behaviors.

LLama-2 can understand intersectional aspects via in-context learning.

Abstract

Amidst the rapid expansion of Machine Learning (ML) and Large Language Models (LLMs), understanding the semantics within their mechanisms is vital. Causal analyses define semantics, while gradient-based methods are essential to eXplainable AI (XAI), interpreting the model's 'black box'. Integrating these, we investigate how a model's mechanisms reveal its causal effect on evidence-based decision-making. Research indicates intersectionality - the combined impact of an individual's demographics - can be framed as an Average Treatment Effect (ATE). This paper demonstrates that hateful meme detection can be viewed as an ATE estimation using intersectionality principles, and summarized gradient-based attention scores highlight distinct behaviors of three Transformer models. We further reveal that LLM Llama-2 can discern the intersectional aspects of the detection through in-context learning and that the learning process could be explained via meta-gradient, a secondary form of gradient. In conclusion, this work furthers the dialogue on Causality and XAI. Our code is available online (see External Resources section).