AI models are not infallible, which is why predictions are often accompanied by confidence scores, and thanks to recent research, these uncertainty estimates have become more accurate, efficient, and scalable.
The researchers Massachusetts Institute of Technology (MIT, Massachusetts, USA) have developed a methodology called IF-COMP to improve the accuracy and efficiency of these uncertainty estimates in machine learning models, increasing their usefulness to researchers and clinicians.
Machine learning is becoming increasingly common in the life sciences, from predicting certain behaviors in animals to analyzing medical images to identify diseases. For such scientific and medical purposes, it is essential for users to know the uncertainty estimate of a particular output or prediction.
If a model identifies pleural effusion in a medical image with 49% confidence, then the model is expected to be correct 49% of the time. However, these estimates are only useful if they are accurate. There is an effort to make these uncertainty estimates more accurate and efficient, as well as applicable to large-scale deep learning models used in safety-critical situations. Not all end users are machine learning experts, so the better the information output, the more informed decisions they can make.
“It's easy to see that these models perform extremely well in some scenarios, and we can assume they'll perform just as well in other scenarios. That makes it particularly important to advance this kind of research to better calibrate the uncertainty in these models and make them more consistent with human notions of uncertainty,” said lead author Nathan Ng.
Traditional uncertainty quantification methods require complex calculations and can be difficult to scale to millions of parameters in machine learning models. They also add complexity and require users to make assumptions not only about the model itself, but also about the data used to train the model.
To avoid users having to make guesses about which models to use, the research team made use of the Minimum Description Length principle (MDL), which is used to better quantify and accommodate uncertainty in model testing iterations.
Can AI predict the onset of autoimmune diseases?
A new study details how an AI algorithm has been trained to identify genes linked to autoimmune diseases, allowing for earlier intervention.
MDL considers all possible labels that the model could assign to a given test point and calculates an uncertainty estimate based on the number of alternative labels that the model did not choose. Alternative labels fit the test point better, so the model is less confident. Each data point is labeled with a certain amount of code, called the probabilistic data complexity. If the model is confident in its prediction, the code will be shorter, but as uncertainty increases, the code length will increase. However, because MDL must evaluate all labels for each data point, it can be very slow and require a lot of computing power.
Therefore, the team developed IF-COMP to make MDL fast enough for use in large-scale deep learning models deployed in clinical settings. IF-COMP is an accurate approximation method that can estimate the complexity of probabilistic data using influence functions. In parallel with this function, the team utilized temperature scaling, a statistical method that improves model output calibration. Combining influence functions and scaling allows the complexity of probabilistic data to be accurately approximated, efficiently generating uncertainty quantification that reflects the true reliability of the model.
Tests have demonstrated that IF-COMP quantifies uncertainty faster and more accurately than other methods. It is also more versatile because it can be applied to a variety of machine learning models. In the future, the researchers hope to apply IF-COMP to larger language models and explore other uses of MDL.
