Researchers at North Carolina State University have developed a novel AI technique to overcome spurious correlations, even if they are unknown. This breakthrough could significantly enhance AI model performance and reliability.
Researchers at North Carolina State University have unveiled a new technique to address one of the most persistent problems in artificial intelligence (AI) model training — spurious correlations. These correlations often cause AI systems to make decisions based on irrelevant or misleading information, hampering the accuracy and reliability of these models.
“This technique is novel in that it can be used even when you have no idea what spurious correlations the AI is relying on,” corresponding author Jung-Eun Kim, an assistant professor of computer science at NC State, said in a news release.
AI models, during their training phase, sometimes latch onto unimportant features due to what is known as simplicity bias. For example, an AI trained to identify dogs in photographs might use collars as the main identifying feature if many training images depict dogs with collars. This can lead to erroneous results, such as incorrectly identifying cats with collars as dogs.
Conventional methods to counteract this problem involve identifying and adjusting the spurious features within the training data. However, in many cases, pinpointing these spurious features is neither straightforward nor even possible, rendering traditional approaches ineffective.
“Our goal with this work was to develop a technique that allows us to sever spurious correlations even when we know nothing about those spurious features,” Kim added.
The new method, termed “data pruning,” involves removing a small subset of the most difficult samples from the training data. These samples typically force the AI model to rely on irrelevant information that leads to spurious correlations.
“There can be significant variation in the data samples included in training data sets,” added Kim. “Some of the samples can be very simple, while others may be very complex. And we can measure how ‘difficult’ each sample is based on how the model behaved during training.”
The hypothesis underlying this approach is that eliminating a small fraction of the most challenging data samples also removes those with spurious features, hence improving model performance without causing significant adverse effects.
The researchers demonstrated that this novel technique produces state-of-the-art results, even outperforming existing methods that require identification of spurious features.
The peer-reviewed paper, titled “Severing Spurious Correlations with Data Pruning,” will be presented at the International Conference on Learning Representations (ICLR) to be held in Singapore from April 24 to 28.
This breakthrough holds significant potential for the field of AI, promising greater accuracy and reliability in AI model outcomes without the exhaustive need to identify and correct spurious correlations individually.
Source: North Carolina State University