GitLab

Machine-learning designs can fail when they attempt to make predictions for individuals who were underrepresented in the datasets they were trained on.

For instance, a design that anticipates the best treatment choice for somebody with a chronic disease may be trained using a dataset that contains mainly male clients. That model may make incorrect predictions for female clients when deployed in a hospital.

To improve results, engineers can attempt stabilizing the training dataset by eliminating information points until all subgroups are represented equally. While dataset balancing is appealing, it often requires getting rid of big amount of data, harming the model's general efficiency.

MIT researchers established a new technique that determines and eliminates 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 enhancing its performance relating to underrepresented groups.

In addition, the strategy can identify hidden sources of predisposition in a training dataset that does not have labels. Unlabeled data are even more widespread than identified information for many applications.

This method might also be combined with other approaches to enhance the fairness of machine-learning models deployed in high-stakes situations. For example, it may one day underrepresented patients aren't misdiagnosed due to a biased AI design.

"Many other algorithms that attempt to address this problem assume each datapoint matters as much as every other datapoint. In this paper, we are revealing that assumption is not real. There are specific points in our dataset that are contributing to this predisposition, and we can discover those data points, remove them, and improve performance," says Kimia Hamidieh, an electrical engineering and computer technology (EECS) graduate trainee at MIT and co-lead author of a paper on this method.

She wrote the paper with co-lead authors Saachi Jain PhD '24 and fellow EECS graduate trainee Kristian Georgiev; Andrew Ilyas MEng '18, PhD '23, a Stein Fellow at Stanford University; and senior authors Marzyeh Ghassemi, an associate teacher in EECS and a member of the Institute of Medical Engineering Sciences and the Laboratory for Details and Decision Systems, and Aleksander Madry, the Cadence Design Systems Professor at MIT. The research will be presented at the Conference on Neural Details Processing Systems.

Removing bad examples

Often, machine-learning models are trained using big datasets gathered from many sources throughout the web. These datasets are far too big to be carefully curated by hand, so they may contain bad examples that hurt design efficiency.

Scientists likewise understand that some information points impact a design's efficiency on certain downstream tasks more than others.

The MIT researchers integrated these two ideas into a method that identifies and gets rid of these bothersome datapoints. They seek to resolve an issue referred to as worst-group mistake, which takes place when a model underperforms on minority subgroups in a training dataset.

The researchers' brand-new method is driven by prior operate in which they presented an approach, called TRAK, that identifies the most crucial training examples for a particular design output.

For this new method, they take inaccurate predictions the design made about minority subgroups and use TRAK to identify which training examples contributed the most to that inaccurate prediction.

"By aggregating this details across bad test predictions in the ideal method, we have the ability to discover the specific parts of the training that are driving worst-group precision down overall," Ilyas explains.

Then they get rid of those particular samples and retrain the design on the remaining information.

Since having more data normally yields much better overall performance, getting rid of just the samples that drive worst-group failures maintains the model's general accuracy while increasing its performance on minority subgroups.

A more available technique

Across three machine-learning datasets, their technique exceeded multiple strategies. In one circumstances, it boosted worst-group accuracy while eliminating about 20,000 less training samples than a standard information balancing technique. Their technique also attained greater accuracy than techniques that require making modifications to the inner functions of a model.

Because the MIT technique involves changing a dataset rather, it would be much easier for a practitioner to use and can be used to numerous kinds of designs.

It can also be utilized when predisposition is unidentified because subgroups in a training dataset are not labeled. By determining 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 anybody can utilize when they are training a machine-learning design. They can take a look at those datapoints and see whether they are aligned with the ability they are trying to teach the design," states Hamidieh.

Using the method to find unknown subgroup predisposition would need intuition about which groups to search for, so the researchers hope to validate it and explore it more totally through future human research studies.

They likewise desire to enhance the performance and reliability of their technique and guarantee the method is available and coastalplainplants.org user friendly for specialists who might someday release it in real-world environments.

"When you have tools that let you critically take a look at the data and find out which datapoints are going to lead to bias or other undesirable behavior, it provides you a primary step towards building designs that are going to be more fair and more trustworthy," Ilyas states.

This work is funded, in part, by the National Science Foundation and the U.S. Defense Advanced Research Projects Agency.