In recent years, there has been a growing focus on ensuring fairness in machine learning (ML) systems. However, this task is complex due to the interplay between fairness and efficiency in deep learning models. This article surveys recent progress in addressing these challenges through various approaches, including model compression techniques, differential privacy, and fairness-aware optimization methods.
Model Compression Techniques
To improve efficiency while maintaining fairness, researchers have explored various model compression techniques such as pruning and quantization (Ahia et al., 2021; Tran et al., 2022). These techniques can reduce the computational cost of ML models without compromising their accuracy. However, they may also affect fairness, as some groups may be underrepresented in the compressed model.
Differential Privacy
Another approach to addressing fairness and privacy is through differential privacy (Dwork et al., 2006). This algorithmic property provides a way to protect sensitive data by adding noise to the model’s outputs, making it difficult to infer information about individual users. However, this can also lead to a trade-off between fairness and accuracy in ML models.
Fairness-Aware Optimization Methods
Recently, there has been growing interest in developing fairness-aware optimization methods that can optimize both efficiency and fairness simultaneously (Tran et al., 2021a; Cummings et al., 2019). These methods aim to find a balance between the two objectives by incorporating fairness constraints into the optimization process.
In summary, ensuring fairness in deep learning models is a complex task that requires careful consideration of both efficiency and fairness. By exploring various approaches, including model compression techniques, differential privacy, and fairness-aware optimization methods, researchers can develop more equitable ML systems that promote better outcomes for all individuals.
