Unpacking Bias Transfer: How Knowledge Distillation Affects Model Fairness

TLDR: This research paper investigates how Knowledge Distillation (KD), a technique for transferring knowledge from large ‘teacher’ models to smaller ‘student’ models, impacts the student’s ability to mitigate biases. The study finds that KD often undermines debiasing capabilities, doesn’t necessarily improve bias mitigation compared to standard training, and that larger teachers don’t guarantee less biased students. The effectiveness of transfer depends on model scale similarity and internal attention patterns. To address these issues, the authors propose three solutions: data augmentation, iterative knowledge distillation, and initializing student models with teacher weights, all of which improve the distillability of debiasing methods.

Machine learning models are incredibly powerful, but they often pick up on hidden biases or “shortcuts” present in the data they are trained on. This can lead to excellent performance on familiar data but significant failures when encountering new, slightly different information. A new research paper, “Do Students Debias Like Teachers? On the Distillability of Bias Mitigation Methods” by Jiali Cheng, Chirag Agarwal, and Hadi Amiri, delves into a critical question: how does a popular technique called Knowledge Distillation (KD) affect a model’s ability to overcome these biases?

Knowledge Distillation is a method typically used to compress large, complex “teacher” models into smaller, more efficient “student” models while retaining much of their original knowledge. However, its impact on a model’s robustness against spurious correlations – those misleading patterns that cause bias – has been largely unexplored, especially at scale.

Unpacking the Research Questions

The researchers set out to answer three main questions:

• To what extent can knowledge distillation transfer debiasing capabilities between models?
• Can knowledge distillation train less biased models compared to standard training?
• What internal mechanisms cause the debiasing behavior to change after distillation?

Key Findings on Bias and Distillation

Through extensive experiments on both natural language inference (NLI) and image classification tasks, the study revealed several important insights:

• Debiasing Capability is Undermined: Overall, the ability of a model to mitigate bias is often weakened after knowledge distillation. Student models tend to become more susceptible to spurious correlations than their teacher counterparts.
• No Benefit from Teacher Knowledge for Debiasing: Surprisingly, training a debiased model doesn’t necessarily improve by injecting knowledge from a teacher model through KD. In many cases, models trained from scratch without KD showed better debiasing performance.
• Variations Across Bias Types: While a model’s overall robustness might seem stable after distillation, there can be significant differences in how it handles various types of biases.
• Internal Mechanisms at Play: The study pinpointed that changes in internal attention patterns and computational circuits within the models are responsible for the distinct behavior observed after KD, particularly when dealing with out-of-distribution data.
• Teacher Size Doesn’t Guarantee Robustness: A larger, more capable teacher model does not automatically lead to a less biased student. The student’s capacity and the complexity of combining debiasing objectives with KD can sometimes hinder effective learning.
• Scale Similarity Matters: Debiasing capabilities transfer most effectively when the teacher and student models are similar in size and complexity.

Also Read:

• SpecKD: A Smarter Way to Distill Knowledge into Smaller Language Models
• New Methods Tackle Contextual Bias in Large Language Models Through Logit Interventions

Proposed Solutions for Better Distillability

Given these findings, the researchers proposed three effective strategies to improve how debiasing methods can be distilled:

1. Data Augmentation (DA): Providing high-quality, debiased data for training significantly improves the distillation process. This addresses the root cause of spurious correlations directly.
2. Iterative Knowledge Distillation (IKD): Instead of a single large jump, knowledge is transferred gradually. A large teacher distills to a slightly smaller student, which then becomes a teacher for an even smaller model, and so on. This step-wise approach ensures smoother knowledge transfer.
3. Initialize Student with Teacher Weights (Init): Starting the student model with weights from the teacher model can give it a head start, helping it to acquire stronger debiasing capabilities and overcome optimization challenges.

The study found that all three solutions improved distillability, with data augmentation yielding the most significant gains. This research is a crucial step towards understanding how knowledge distillation impacts bias mitigation and provides valuable insights for designing more robust and fair machine learning models in the future.