Model Card for GANdhi

Model description

GANdhi is a Generative Adversarial Network (GAN) based model designed to score the fairness of a dataset distribution. The model helps assess biases in datasets and provides valuable feedback to researchers and developers, ensuring a more equitable representation of data across various domains.

The page serves as a portflio place for the project, as the training data and model weight involves classfied information.

Intended use

GANdhi is intended for:

• Fairness assessment in dataset distribution
• Bias detection and mitigation
• Evaluating and improving dataset quality
• The model aims to assist in creating datasets that are more balanced and representative, leading to better-performing machine learning models that are less biased.

Training data

The GANdhi model was trained on a diverse set of datasets, which were preprocessed and curated to cover a broad range of domains and subjects. The exact datasets used in the training process are proprietary information.

Model performance

GANdhi demonstrated effective performance in identifying biases and assessing fairness across various dataset distributions:

• Fairness scoring accuracy: 91.4% mean on ImageNet 1k
• Bias detection rate: 93.61% mean on COCO validation set

Limitations

The GANdhi model has the following limitations:

• Sensitivity to noise: The model may be sensitive to noise present in the dataset, which could impact its fairness assessment performance.
• Incomplete domain coverage: While the model has been trained on diverse datasets, it may not cover all possible domains and contexts, leading to potential discrepancies in its performance.