MIT Researchers Develop New Technique to Reduce AI Bias While Maintaining Accuracy

Researchers reduce bias in AI models while preserving or improving accuracy

Machine-learning models can fail when they try to make predictions for individuals who were underrepresented in the datasets they were trained on. For instance, a model that predicts the best treatment option for someone with a chronic disease may be trained using a dataset that contains mostly male patients. That model might make incorrect predictions for female patients when deployed in a hospital. To improve outcomes, engineers can try balancing the training dataset by removing data points until all subgroups are represented equally. While dataset balancing is promising, it often requires removing large amount of data, hurting the model's overall performance. MIT researchers developed a new technique that identifies and removes specific points in a training dataset that contribute most to a model's failures on minority subgroups. By removing far fewer datapoints than other approaches, this technique maintains the overall accuracy of the model while improving its performance regarding underrepresented groups. In addition, the technique can identify hidden sources of bias in a training dataset that lacks labels. Unlabeled data are far more prevalent than labeled data for many applications. This method could also be combined with other approaches to improve the fairness of machine-learning models deployed in high-stakes situations. For example, it might someday help ensure underrepresented patients aren't misdiagnosed due to a biased AI model. "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. There are specific points in our dataset that are contributing to this bias, and we can find those data points, remove them, and get better performance," says Kimia Hamidieh, an electrical engineering and computer science (EECS) graduate student at MIT and co-lead author of a paper on this technique. She wrote the paper with co-lead authors Saachi Jain PhD '24 and fellow EECS graduate student Kristian Georgiev; Andrew Ilyas MEng '18, PhD '23, a Stein Fellow at Stanford University; and senior authors Marzyeh Ghassemi, an associate professor in EECS and a member of the Institute of Medical Engineering Sciences and the Laboratory for Information and Decision Systems, and Aleksander Madry, the Cadence Design Systems Professor at MIT. The research will be presented at the Conference on Neural Information Processing Systems. Removing bad examples Often, machine-learning models are trained using huge datasets gathered from many sources across the internet. These datasets are far too large to be carefully curated by hand, so they may contain bad examples that hurt model performance. Scientists also know that some data points impact a model's performance on certain downstream tasks more than others. The MIT researchers combined these two ideas into an approach that identifies and removes these problematic datapoints. They seek to solve a problem known as worst-group error, which occurs when a model underperforms on minority subgroups in a training dataset. The researchers' new technique is driven by prior work in which they introduced a method, called TRAK, that identifies the most important training examples for a specific model output. For this new technique, they take incorrect predictions the model made about minority subgroups and use TRAK to identify which training examples contributed the most to that incorrect prediction. "By aggregating this information across bad test predictions in the right way, we are able to find the specific parts of the training that are driving worst-group accuracy down overall," Ilyas explains. Then they remove those specific samples and retrain the model on the remaining data. Since having more data usually yields better overall performance, removing just the samples that drive worst-group failures maintains the model's overall accuracy while boosting its performance on minority subgroups. A more accessible approach Across three machine-learning datasets, their method outperformed multiple techniques. In one instance, it boosted worst-group accuracy while removing about 20,000 fewer training samples than a conventional data balancing method. Their technique also achieved higher accuracy than methods that require making changes to the inner workings of a model. Because the MIT method involves changing a dataset instead, it would be easier for a practitioner to use and can be applied to many types of models. It can also be utilized when bias is unknown because subgroups in a training dataset are not labeled. By identifying datapoints that contribute most to a feature the model is learning, they can understand the variables it is using to make a prediction. "This is a tool anyone can use when they are training a machine-learning model. They can look at those datapoints and see whether they are aligned with the capability they are trying to teach the model," says Hamidieh. Using the technique to detect unknown subgroup bias would require intuition about which groups to look for, so the researchers hope to validate it and explore it more fully through future human studies. They also want to improve the performance and reliability of their technique and ensure the method is accessible and easy-to-use for practitioners who could someday deploy it in real-world environments. "When you have tools that let you critically look at the data and figure out which datapoints are going to lead to bias or other undesirable behavior, it gives you a first step toward building models that are going to be more fair and more reliable," Ilyas says. This work is funded, in part, by the National Science Foundation and the U.S. Defense Advanced Research Projects Agency.

Researchers reduce bias in AI models while preserving or improving accuracy

Machine-learning models can fail when they try to make predictions for individuals who were underrepresented in the datasets they were trained on. For instance, a model that predicts the best treatment option for someone with a chronic disease may be trained using a dataset that contains mostly male patients. That model might make incorrect predictions for female patients when deployed in a hospital. To improve outcomes, engineers can try balancing the training dataset by removing data points until all subgroups are represented equally. While dataset balancing is promising, it often requires removing large amount of data, hurting the model's overall performance. MIT researchers developed a new technique that identifies and removes specific points in a training dataset that contribute most to a model's failures on minority subgroups. By removing far fewer datapoints than other approaches, this technique maintains the overall accuracy of the model while improving its performance regarding underrepresented groups. In addition, the technique can identify hidden sources of bias in a training dataset that lacks labels. Unlabeled data are far more prevalent than labeled data for many applications. This method could also be combined with other approaches to improve the fairness of machine-learning models deployed in high-stakes situations. For example, it might someday help ensure underrepresented patients aren't misdiagnosed due to a biased AI model. "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. There are specific points in our dataset that are contributing to this bias, and we can find those data points, remove them, and get better performance," says Kimia Hamidieh, an electrical engineering and computer science (EECS) graduate student at MIT and co-lead author of a paper on this technique. She wrote the paper with co-lead authors Saachi Jain PhD '24 and fellow EECS graduate student Kristian Georgiev; Andrew Ilyas MEng '18, PhD '23, a Stein Fellow at Stanford University; and senior authors Marzyeh Ghassemi, an associate professor in EECS and a member of the Institute of Medical Engineering Sciences and the Laboratory for Information and Decision Systems, and Aleksander Madry, the Cadence Design Systems Professor at MIT. The research will be presented at the Conference on Neural Information Processing Systems. Removing bad examples Often, machine-learning models are trained using huge datasets gathered from many sources across the internet. These datasets are far too large to be carefully curated by hand, so they may contain bad examples that hurt model performance. Scientists also know that some data points impact a model's performance on certain downstream tasks more than others. The MIT researchers combined these two ideas into an approach that identifies and removes these problematic datapoints. They seek to solve a problem known as worst-group error, which occurs when a model underperforms on minority subgroups in a training dataset. The researchers' new technique is driven by prior work in which they introduced a method, called TRAK, that identifies the most important training examples for a specific model output. For this new technique, they take incorrect predictions the model made about minority subgroups and use TRAK to identify which training examples contributed the most to that incorrect prediction. "By aggregating this information across bad test predictions in the right way, we are able to find the specific parts of the training that are driving worst-group accuracy down overall," Ilyas explains. Then they remove those specific samples and retrain the model on the remaining data. Since having more data usually yields better overall performance, removing just the samples that drive worst-group failures maintains the model's overall accuracy while boosting its performance on minority subgroups. A more accessible approach Across three machine-learning datasets, their method outperformed multiple techniques. In one instance, it boosted worst-group accuracy while removing about 20,000 fewer training samples than a conventional data balancing method. Their technique also achieved higher accuracy than methods that require making changes to the inner workings of a model. Because the MIT method involves changing a dataset instead, it would be easier for a practitioner to use and can be applied to many types of models. It can also be utilized when bias is unknown because subgroups in a training dataset are not labeled. By identifying datapoints that contribute most to a feature the model is learning, they can understand the variables it is using to make a prediction. "This is a tool anyone can use when they are training a machine-learning model. They can look at those datapoints and see whether they are aligned with the capability they are trying to teach the model," says Hamidieh. Using the technique to detect unknown subgroup bias would require intuition about which groups to look for, so the researchers hope to validate it and explore it more fully through future human studies. They also want to improve the performance and reliability of their technique and ensure the method is accessible and easy-to-use for practitioners who could someday deploy it in real-world environments. "When you have tools that let you critically look at the data and figure out which datapoints are going to lead to bias or other undesirable behavior, it gives you a first step toward building models that are going to be more fair and more reliable," Ilyas says. This work is funded, in part, by the National Science Foundation and the U.S. Defense Advanced Research Projects Agency.

New technique reduces bias in AI models while preserving or improving accuracy

Machine-learning models can fail when they try to make predictions for individuals who were underrepresented in the datasets they were trained on. For instance, a model that predicts the best treatment option for someone with a chronic disease may be trained using a dataset that contains mostly male patients. That model might make incorrect predictions for female patients when deployed in a hospital. To improve outcomes, engineers can try balancing the training dataset by removing data points until all subgroups are represented equally. While dataset balancing is promising, it often requires removing large amounts of data, hurting the model's overall performance. MIT researchers have developed a new technique that identifies and removes specific points in a training dataset that contribute most to a model's failures on minority subgroups. By removing far fewer data points than other approaches, this technique maintains the overall accuracy of the model while improving its performance regarding underrepresented groups. In addition, the technique can identify hidden sources of bias in a training dataset that lacks labels. Unlabeled data are far more prevalent than labeled data for many applications. This method could also be combined with other approaches to improve the fairness of machine-learning models deployed in high-stakes situations. For example, it might someday help ensure underrepresented patients aren't misdiagnosed due to a biased AI model. "Many other algorithms that try to address this issue assume each data point matters as much as every other data point. In this paper, we are showing that assumption is not true. "There are specific points in our dataset that are contributing to this bias, and we can find those data points, remove them, and get better performance," says Kimia Hamidieh, an electrical engineering and computer science (EECS) graduate student at MIT and co-lead author of a paper on this technique posted to the arXiv preprint server. She wrote the paper with co-lead authors Saachi Jain Ph.D. '24 and fellow EECS graduate student Kristian Georgiev; Andrew Ilyas MEng '18, Ph.D. '23, a Stein Fellow at Stanford University; and senior authors Marzyeh Ghassemi, an associate professor in EECS and a member of the Institute of Medical Engineering Sciences and the Laboratory for Information and Decision Systems, and Aleksander Madry, the Cadence Design Systems Professor at MIT. The research will be presented at the Conference on Neural Information Processing Systems (NeurIPS 2024), held in Vancouver Dec. 10-15. Removing bad examples Often, machine-learning models are trained using huge datasets gathered from many sources across the internet. These datasets are far too large to be carefully curated by hand, so they may contain bad examples that hurt model performance. Scientists also know that some data points impact a model's performance on certain downstream tasks more than others. The MIT researchers combined these two ideas into an approach that identifies and removes these problematic data points. They seek to solve a problem known as worst-group error, which occurs when a model underperforms on minority subgroups in a training dataset. The researchers' new technique is driven by prior work in which they introduced a method, called TRAK, that identifies the most important training examples for a specific model output. For this new technique, they take incorrect predictions the model made about minority subgroups and use TRAK to identify which training examples contributed the most to that incorrect prediction. "By aggregating this information across bad test predictions in the right way, we are able to find the specific parts of the training that are driving worst-group accuracy down overall," Ilyas explains. Then they remove those specific samples and retrain the model on the remaining data. Since having more data usually yields better overall performance, removing just the samples that drive worst-group failures maintains the model's overall accuracy while boosting its performance on minority subgroups. A more accessible approach Across three machine-learning datasets, their method outperformed multiple techniques. In one instance, it boosted worst-group accuracy while removing about 20,000 fewer training samples than a conventional data balancing method. Their technique also achieved higher accuracy than methods that require making changes to the inner workings of a model. Because the MIT method involves changing a dataset instead, it would be easier for a practitioner to use and can be applied to many types of models. It can also be utilized when bias is unknown, because subgroups in a training dataset are not labeled. By identifying data points that contribute most to a feature the model is learning, they can understand the variables it is using to make a prediction. "This is a tool anyone can use when they are training a machine-learning model. They can look at those data points and see whether they are aligned with the capability they are trying to teach the model," says Hamidieh. Using the technique to detect unknown subgroup bias would require intuition about which groups to look for, so the researchers hope to validate it and explore it more fully through future human studies. They also want to improve the performance and reliability of their technique and ensure the method is accessible and easy-to-use for practitioners who could someday deploy it in real-world environments. "When you have tools that let you critically look at the data and figure out which data points are going to lead to bias or other undesirable behavior, it gives you a first step toward building models that are going to be more fair and more reliable," Ilyas says.