BAdd: Bias Mitigation through Bias Addition
Ioannis Sarridis, Christos Koutlis, Symeon Papadopoulos, Christos Diou
TL;DR
BAdd is a straightforward bias mitigation technique that enhances fairness in computer vision models by integrating attribute features into the backbone, outperforming existing methods especially in complex multi-attribute bias scenarios.
Contribution
The paper introduces BAdd, a novel bias mitigation method that effectively handles multi-attribute biases in CV datasets, surpassing state-of-the-art performance.
Findings
Achieves +27.5% accuracy on FB-Biased-MNIST.
Achieves +5.5% accuracy on CelebA.
Outperforms existing bias mitigation methods on seven benchmarks.
Abstract
Computer vision (CV) datasets often exhibit biases that are perpetuated by deep learning models. While recent efforts aim to mitigate these biases and foster fair representations, they fail in complex real-world scenarios. In particular, existing methods excel in controlled experiments involving benchmarks with single-attribute injected biases, but struggle with multi-attribute biases being present in well-established CV datasets. Here, we introduce BAdd, a simple yet effective method that allows for learning fair representations invariant to the attributes introducing bias by incorporating features representing these attributes into the backbone. BAdd is evaluated on seven benchmarks and exhibits competitive performance, surpassing state-of-the-art methods on both single- and multi-attribute benchmarks. Notably, BAdd achieves +27.5% and +5.5% absolute accuracy improvements on the…
Peer Reviews
Decision·Submitted to ICLR 2025
- The paper is clear and well-written. - The core idea seems interesting and is supported by various experiments.
The idea of Bias Injection to Mitigate Bias has been used in the following work [1]. Although the idea is that a bias injection module, can prevent the loss spike. However, when the loss is forced to be zero, it needs to overcorrect the bias injection module, does it lead to correct features? For the bias-aligned examples, the network can probably take the shortcut. Hence, the learning needs to happen just which Bias Corrected samples. In case B_c >> B_a won’t it affect the learning of diverse
1.BAdd reduces bias effectively by adding bias-related features into the training, helping the model avoid being influenced by biased data. 2.The authors showed that BAdd works well on different datasets, with consistent improvements in various bias situations, proving that the method is scalable and works in different applications.
1.BAdd requires a classifier or labels that identify the bias, which may not always be available. The authors could consider ways to detect and handle biases automatically without needing predefined labels. 2.The method mainly addresses visual biases, and it’s unclear if it works for other types of biases, like those in text, limiting its use beyond visual data. 3.The paper is clear, but lacks a detailed comparison to standard deep learning training. It’s not explained how the bias-detecting cla
1- The problem is described clearly. 2- The investigated problem is important.
1- The BAdd approach seems similar to FLAC [1]. 2- On page 3, The author claims that BAdd is easily applied to any network architecture and to any CV dataset. This claim needs to be proven. 3- The paper does not discuss the computational complexity or scalability of the proposed approach in detail, which could be a concern for large-scale applications. 4- Limited Ablation Studies: While the paper includes some ablation studies, more extensive ablations could strengthen the claims about the i
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdversarial Robustness in Machine Learning