In a collaborative effort between an Oregon State University doctoral student and researchers at Adobe, a groundbreaking technique known as FairDeDup has been developed to address the issue of social biases in artificial intelligence systems. The new method, FairDeDup, aims to make AI systems less socially biased by implementing a cost-effective approach to training processes.
FairDeDup, short for fair deduplication, focuses on removing redundant information from the data used to train AI systems. This process, known as deduplication, not only reduces the high computing costs associated with training AI but also helps mitigate biases present in the datasets. The datasets, often collected from the internet, reflect biases prevalent in society. When these biases are encoded in AI models, they can perpetuate unfair ideas and behaviors.
By incorporating fairness considerations into the deduplication process, FairDeDup aims to understand how bias prevalence can be mitigated. One of the key challenges highlighted by the researchers is the risk of AI systems reinforcing harmful social biases. For example, an AI system might automatically display images of only white men when asked to show a picture of a CEO or doctor. The intended use case, however, may be to show a diverse representation of people.
FairDeDup operates by thinning out datasets of image captions sourced from the web using a technique called pruning. Pruning involves selecting a subset of data that accurately represents the entire dataset. By pruning the data in a content-aware manner, FairDeDup enables informed decision-making about which data points to retain and which ones to discard. This approach ensures that the training process is both cost-effective and accurate, while also promoting fairness.
Apart from biases related to occupation, race, and gender, FairDeDup also targets biases tied to age, geography, and culture. Addressing these biases during the dataset pruning phase is crucial for creating AI systems that uphold social justice. The researchers emphasize that their work does not impose a singular notion of fairness on AI systems. Instead, it provides a framework for defining fairness based on specific settings and user preferences.
The development of FairDeDup involved collaboration between Eric Slyman, a doctoral student at Oregon State University, and researchers including Stefan Lee, an assistant professor at OSU, as well as Scott Cohen and Kushal Kafle from Adobe. By pooling their expertise, the team was able to refine the FairDeDup algorithm and demonstrate its effectiveness in promoting fairness in AI systems.
Overall, the FairDeDup algorithm represents a significant step towards addressing social biases in AI training. By integrating fairness considerations into the deduplication process, this innovative technique offers a promising solution to the challenge of bias prevalence in artificial intelligence systems.
Leave a Reply