New Technique Overcomes Spurious Correlations In AI Models

North Carolina State University researchers have developed a novel technique to address spurious correlations in AI models by identifying and removing problematic data samples from training datasets. This method does not require prior knowledge of the specific spurious features causing issues, making it applicable even when performance problems arise without clear explanations. The approach focuses on eliminating small portions of difficult-to-interpret training data that contribute to misleading correlations, thereby improving model reliability. The findings will be presented at the International Conference on Learning Representations (ICLR) in Singapore from April 24-28, 2025.

New Technique Addresses Spurious Correlations in AI

AI models often rely on spurious correlations, making decisions based on unimportant and potentially misleading information. Researchers have now discovered these learned spurious correlations can be traced to a very small subset of the training data and have demonstrated a technique that overcomes the problem. This technique is novel in that it can be used even when you have no idea what spurious correlations the AI is relying on, says Jung-En Kim, corresponding author of a paper on the work and an assistant professor of computer science at North Carolina State University.

Spurious correlations are generally caused by simplicity bias during training, where models prioritize easier-to-detect features over more meaningful ones. For example, an AI might identify dogs by their collars instead of more relevant features like ears or fur because collars are simpler to detect. This reliance on spurious correlations can lead to poor performance when deployed in real-world scenarios where such patterns may not hold.

Conventional techniques to address this issue require prior knowledge of the problematic features, limiting their applicability. However, a novel approach has been developed that measures the difficulty of each data sample during training and removes those that contribute to reliance on irrelevant features. By focusing on the subset of data most responsible for spurious learning, it achieves targeted improvements in model robustness.

This new method does not require prior knowledge of spurious features and can be applied even when the cause of the problem is unclear. It effectively balances accuracy and generalization, offering a versatile solution to reduce simplicity bias and enhance model reliability in real-world applications. The technique has demonstrated state-of-the-art results compared to previous work, effectively eliminating spurious correlations without significantly harming performance.

The approach represents a significant advancement in addressing simplicity bias and improving model generalization. It highlights the importance of careful data curation in AI development and offers a practical solution for building models that perform better in real-world scenarios where training correlations may not hold.