As deep learning models become tasked with more and more decisions that impact human lives, such as criminal recidivism, loan repayment, and face recognition for law enforcement, bias is becoming a growing concern. Debiasing algorithms are typically split into three paradigms: pre-processing, in-processing, and post-processing. However, in computer vision or natural language applications, it is common to start with a large generic model and then fine-tune to a specific use-case. Pre-or in-processing methods would require retraining the entire model from scratch, while post-processing methods only have black-box access to the model, so they do not leverage the weights of the trained model. Creating debiasing algorithms specifically for this fine-tuning use-case has largely been neglected. In this work, we initiate the study of a new paradigm in debiasing research, intra-processing, which sits between in-processing and post-processing methods. Intra-processing methods are designed specifically to debias large models which have been trained on a generic dataset, and fine-tuned on a more specific task. We show how to repurpose existing in-processing methods for this use-case, and we also propose three baseline algorithms: random perturbation, layerwise optimization, and adversarial debiasing. We evaluate these methods across three popular datasets from the AIF360 toolkit, as well as on the CelebA faces dataset.

However, in computer vision or natural language applications, it is common to start with a large generic model and then fine-tune to a specific use-case.

It would be clearer if the accuracy rates of the methods were provided in a table. While this is not true as this method is a model-agnostic post-processing method that only changes the weights of the adversary. So, this model should be studied in the experiments. I also believe the authors still need to run more experiments and compare with more exiting works. Therefore, I will keep my score unchanged.

This work presents a simple idea that works. The proposal can benefit from an improved presentation of the material and a more comprehensive related work section. In particular, prior work such as Multiaccuracy: Black-Box Post-Processing for Fairness in Classification should be cited. Besides, we strongly encourage the authors to incorporate in the final version of the paper the results presented in the rebuttal. Also, we suggest the authors present the results in a standard way that makes comparison with other work easier.

As deep learning models become tasked with more and more decisions that impact human lives, such as criminal recidivism, loan repayment, and face recognition for law enforcement, bias is becoming a growing concern. Debiasing algorithms are typically split into three paradigms: pre-processing, in-processing, and post-processing. However, in computer vision or natural language applications, it is common to start with a large generic model and then fine-tune to a specific use-case. Pre- or in-processing methods would require retraining the entire model from scratch, while post-processing methods only have black-box access to the model, so they do not leverage the weights of the trained model. Creating debiasing algorithms specifically for this fine-tuning use-case has largely been neglected.