# Towards Trustworthy and Aligned Machine Learning: A Data-centric Survey with Causality Perspectives

**Authors:** Haoyang Liu, Maheep Chaudhary, Haohan Wang  
**arXiv:** [2307.16851](https://arxiv.org/abs/2307.16851)  
**Date:** 2023-07-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.16851v1/

## Key points

- Authors Haoyang Liu, Maheep Chaudhary, and Haohan Wang focus on trustworthiness in machine learning.
- Survey covers applications and research areas like robustness, security, interpretability, and fairness.
- Observes convergence of methods across trustworthy machine learning subfields due to Pearl's hierarchy of causality.
- Aims to provide a unified understanding of trustworthy machine learning development by connecting concepts to Pearl's causal hierarchy.
- Explores trustworthiness of large pretrained models with techniques such as fine-tuning, parameter-efficient fine-tuning, prompting, and reinforcement learning with human feedback.
- Connects these techniques to standard empirical risk minimization (ERM) training methods for a principled understanding in large pretrained models.
- Offers a detailed review of existing methods and insights into potential future aspects related to the survey findings.
- Links mathematical vocabulary across robustness, adversarial robustness, interpretability, and fairness to enhance overall field understanding.

## Summary

In their comprehensive survey titled "Towards Trustworthy and Aligned Machine Learning: A Data-centric Survey with Causality Perspectives," authors Haoyang Liu, Maheep Chaudhary, and Haohan Wang delve into the critical topic of trustworthiness in machine learning. The survey covers a wide range of applications and research areas such as robustness, security, interpretability, and fairness, highlighting the advancements made over the last decade to address these challenges. One key observation made by the authors is the convergence of methods developed independently across various trustworthy machine learning subfields. They attribute this convergence to Pearl's hierarchy of causality, which provides a unifying framework for understanding these techniques. By presenting a unified set of concepts and connecting them to Pearl's causal hierarchy, the authors aim to foster a more cohesive understanding of trustworthy machine learning development. The survey also explores the trustworthiness of large pretrained models, summarizing dominant techniques like fine-tuning, parameter-efficient fine-tuning, prompting, and reinforcement learning with human feedback. By drawing connections between these techniques and standard empirical risk minimization (ERM) training methods, the authors extend the principled understanding of trustworthy methods to new approaches in large pretrained models. Furthermore,the authors provide a detailed review of existing methods under this perspective and offer insights into potential future aspects related to their survey findings. By linking mathematical vocabulary across robustness, adversarial robustness, interpretability,and fairness , they aim to enhance the field's overall understanding. Overall,this survey serves as a valuable resource for researchers and practitioners in machine learning seeking to deepen their knowledge of trustworthy methodologies.For more information on this topic , readers are encouraged to visit http://trustai.one.