Downstream Bias

Downstream bias in machine learning refers to the perpetuation or amplification of biases present in training data, leading to unfair or discriminatory outcomes in deployed models. Current research focuses on mitigating this bias through interventions at both the pre-training and fine-tuning stages of large language models (LLMs) and other architectures, exploring techniques like projective methods and parameter-efficient fine-tuning (PEFT). Understanding and addressing downstream bias is crucial for ensuring fairness and trustworthiness in AI systems across various applications, prompting investigations into feature importance disparities and the development of improved bias detection and mitigation strategies.