DISCO: New Method Reduces Bias in Deep Learning Using Conditional Distance Correlation
A new technique named DISCO has been unveiled by researchers to address bias in deep learning models through the use of conditional distance correlation. This method is detailed in a paper titled 'DISCO: Mitigating Bias in Deep Learning with Conditional Distance Correlation,' which introduces the Standard Anti-Causal Model (SAM). This causal framework elucidates the mechanisms behind bias and sets a criterion for conditional independence to ensure causal stability. The authors also created two efficient estimators, DISCO$_m$ and sDISCO, facilitating independence regularization in gradient-based models. Evaluated on six varied datasets, these approaches consistently outperform or equal existing bias reduction methods while needing fewer hyperparameters and accommodating multi-bias situations. The source code is publicly accessible, and the paper is available on arXiv in the Computer Vision and Pattern Recognition section.
Key facts
- DISCO uses conditional distance correlation to mitigate bias in deep learning.
- The Standard Anti-Causal Model (SAM) provides a causal framework for bias mechanisms.
- DISCO$_m$ and sDISCO are efficient estimators of conditional distance correlation.
- The methods were tested on six diverse datasets.
- They outperform or match existing bias mitigation approaches.
- Fewer hyperparameters are required compared to other methods.
- The approach scales to multi-bias scenarios.
- Source code is available at the provided URL.
Entities
Institutions
- arXiv