Recent advancements in artificial intelligence (AI) are shaping the future of machine learning, particularly in addressing the biases that can emerge when these systems are applied to real-world scenarios. Researchers at the Massachusetts Institute of Technology (MIT) have developed an innovative technique aimed at improving the accuracy of AI models for underrepresented groups by eliminating specific problematic data points from training datasets.
The issue of predictive failures among individuals from minority subgroups in machine learning has garnered significant attention. These failures often occur when models are trained on datasets that predominantly feature data from one demographic, which can lead to inappropriate or incorrect predictions. For example, a model developed to determine the best treatment pathways for chronic illnesses might yield unreliable outcomes for patients who are not adequately represented in the training data, such as women if the dataset is male-heavy.
In their research, the MIT team sought to refine the data selection process by creating a method that specifically targets and removes data points contributing to these inaccuracies, without significantly disrupting the overall performance of the model. According to co-lead author Kimia Hamidieh, an electrical engineering and computer science graduate student at MIT, “Many other algorithms that try to address this issue assume each datapoint matters as much as every other datapoint. In this paper, we are showing that assumption is not true."
The innovative approach involves aggregating information from incorrect predictions to identify which training examples lead to the lowest performance for these minority groups. This technique is an evolution of prior work that introduced a method called TRAK, which identifies critical training examples that influence model outputs. By isolating specific articles from the training dataset and removing them, the researchers found they could enhance the model’s performance for underrepresented demographics while maintaining its overall predictive accuracy.
During testing across three separate machine-learning datasets, this new method outperformed various existing techniques. Notably, it was able to enhance worst-group accuracy while removing approximately 20,000 fewer training samples than traditional data balancing methods. Furthermore, since the MIT approach adjusts the dataset rather than altering the underlying model structure, it presents a more flexible option for practitioners in the field.
Beyond mitigating bias, this technique also holds promise for tackling the challenges associated with unlabeled datasets. As unsupervised learning continues to expand, identifying unlabelled group biases may become integral to developing fair AI applications. Hamidieh notes the broad applicability of their method, stating, “This is a tool anyone can use when they are training a machine-learning model."
The research, which will be presented at the Conference on Neural Information Processing Systems, aims to improve the reliability of machine-learning systems in critical areas such as healthcare, where misdiagnoses can have severe ramifications for patients. Senior author Andrew Ilyas highlighted the significance of this tool, suggesting that it provides an essential step towards creating AI models that exhibit both fairness and reliability.
Supported in part by the National Science Foundation and the U.S. Defense Advanced Research Projects Agency, the MIT researchers intend to further validate and optimise their technique through future studies, with the aspiration of making these tools readily available for practitioners developing machine-learning models in various real-world contexts.
Source: Noah Wire Services