Researchers at Seoul National University develop technology to minimize bias in artificial intelligence (AI). The method improves judgment by separately extracting data to minimize false correlations./Courtesy of Chosun DB

Domestic researchers have developed a technology to eliminate biases that artificial intelligence (AI) possesses. As AI enters everyday life, concerns about side effects arising from biases have increased, leading to expectations that this technology will be used for safe AI utilization.

A research team led by Professor Lee Jeong-woo of the Department of Electrical and Computer Engineering at Seoul National University (Hodu AI CEO) announced on the 26th that they have found a learning method to reduce bias in AI.

AI bias is an element that hinders proper judgment, much like human biases. For instance, when AI is tasked with classifying cows and camels, if it is trained on pictures of cows in pastures and camels in deserts, it may classify a photo of a camel in a pasture as a cow. This phenomenon is a representative case of bias, resulting from AI learning both the essential characteristics of the animals and irrelevant peripheral data.

When artificial intelligence (AI) experiments with distinguishing between cows and camels in a photo, bias emerges from the surrounding environment. A typical example of bias is classifying as a camel when there is a cow with a desert background, and classifying as a cow when there is a camel with a pasture background./Courtesy of Seoul National University College of Engineering

As AI becomes more integrated into everyday life, bias is emerging as a social issue. If banks use AI for loan assessments, they may reference supplementary information such as race, gender, and age alongside income and asset levels. This could lead to individuals of specific races and genders with lower income levels being discriminated against, regardless of their abilities.

The research team developed a 'bias reduction technology' that allows AI to make judgments regardless of 'spurious correlations' within the data. The strategy is to identify the key characteristics of the data to minimize the bias in the AI model.

The team first created an AI model characterized by strong bias and calculated the likelihood of spurious correlations in the training data. They then sought combinations that minimized spurious correlations, re-extracted, and retrained to improve accuracy. As a result, the analysis accuracy improved by 21% compared to existing AI models, showing high reliability.

The researchers noted, "We made it possible to extract more training data where the probability of spurious correlations is low," adding that they aim to ensure that the AI model gradually does not rely on spurious correlations.

The research team expects that the technology developed this time can eliminate biases hidden in various forms of data such as images, medical, and legal fields. This technology is used as a core component of the 'Bias Removal AI' engine of Hodu AI, founded by Professor Lee Jeong-woo.

Professor Lee Jeong-woo said, "I will contribute to enhancing the technological level of Korean AI startups by developing innovative technologies."

The research results were presented at the international conference 'Neural Information Processing Systems (NeurIPS) 2024,' held in Vancouver, Canada, from the 9th to the 15th.