Delphi of GANs for Bias Detection in Pre-trained Models

This project implements a Delphi of GANs approach to detect bias in pre-trained language models, specifically BERT and DistilBERT. The technique provides a novel way to assess and compare bias levels in different models across multiple dimensions such as gender, racial, and age bias.

Motivating Article

Di Zio, S., Calleo, Y., & Bolzan, M. (2023). Delphi-based visual scenarios: an innovative use of generative adversarial networks. Futures, 154, 103280.

Results

Concept: Delphi of GANs

The Delphi of GANs approach is inspired by the Delphi method, which involves a panel of experts making predictions or assessments. In our case, each GAN serves as an "expert" trained on a specific type of bias data. By using multiple GANs, we can get a more comprehensive and robust assessment of bias in language models.

Why is this technique useful?

1. Multi-dimensional Bias Detection: It allows for simultaneous assessment of multiple types of bias (e.g., gender, racial, age).
2. Comparative Analysis: We can compare bias levels between different models (e.g., BERT vs. DistilBert).
3. Quantitative Metrics: Provides numerical scores for bias, allowing for objective comparisons.
4. Adaptability: The technique can be extended to other types of bias by training new GANs on relevant datasets.

Research-based Bias Thresholds

We use research-based thresholds to interpret the bias scores:

• Gender Bias: Unbiased ≤ 0.25, Biased ≥ 0.5 (Caliskan et al., 2017)
• Racial Bias: Unbiased ≤ 0.3, Biased ≥ 0.6 (Manzini et al., 2019)
• Age Bias: Unbiased ≤ 0.2, Biased ≥ 0.45 (Díaz et al., 2018) These thresholds help categorize models as unbiased, moderately biased, or significantly biased.

Process

1. Data Preparation:
   • Download and preprocess datasets for gender, racial, and age bias.
2. GAN Creation and Training:
   • For each bias type:
     • Create a GAN with a generator and discriminator.
     • Train the GAN on the corresponding bias dataset.
3. Bias Detection:
   • Load pre-trained models (BERT and DistilBERT).
   • For each model:
     • Generate predictions on test sentences.
     • Use trained GANs to evaluate these predictions.
     • Calculate bias scores based on GAN discriminator outputs.
4. Interpretation:
   • Compare bias scores against research-based thresholds.
   • Provide an overall bias assessment and detailed explanations.

Examples of Bias Found with Pre-trained Models

Classifier Bias Detected: BERT Model

BERT (Bidirectional Encoder Representations from Transformers) is an example of a bias pre-trained model. Its Transformer architecture and self-attention mechanisms allow it to capture complex relationships between words, making it more flexible and less biased towards specific patterns. BERT's pre-training on a large and diverse corpus of text data also contributes to its ability to generalize well to various downstream tasks.

Classifier Bias Detected: DistilBERT Model

DistilBERT is a distilled version of BERT that aims to reduce the model size while maintaining performance. However, the distillation process may introduce additional biases. DistilBERT has been selected as an example of a bias model in this project to demonstrate the capability of the Delphi of GANs approach in detecting and quantifying bias in different pre-trained models.

Usefulness in Assessing Pre-trained Model Classifiers

Understanding bias in pre-trained models is crucial for several reasons:

1. Ethical AI: Helps in developing more fair and unbiased AI systems.
2. Model Selection: Aids in choosing appropriate models for specific tasks, considering bias implications.
3. Transparency: Provides insights into model behavior, crucial for responsible AI deployment.
4. Improvement Opportunities: Identifies areas where models can be fine-tuned or debiased.
5. Regulatory Compliance: Assists in meeting emerging AI regulations and ethical guidelines.

Disclaimer

This code is for research and educational purposes only.

License

Copyright 2024 Eric Yocam

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.