Hey guys, it has been a moment since I last posted on my blog. It has been a busy year with school picking up and my summer internship. But lucky for you, I am back — hoping to start blogging regularly again! Let’s get straight into it.
Machine-learning models often provide predictions with an accompanying confidence level, essential in high-stake areas like medical imaging or job application filtering. However, the utility of these confidence estimates hinges on their accuracy. For instance, a model claiming 49% confidence in diagnosing pleural effusion should be correct approximately 49% of the time.
MIT researchers have developed a novel approach that enhances the accuracy and efficiency of these uncertainty estimates. Their method not only surpasses existing techniques in precision but is also scalable, making it applicable to the vast deep-learning models used in critical domains such as healthcare.
This innovation equips users, including those without in-depth machine-learning knowledge, with more reliable information, aiding decisions on whether to trust a model’s predictions or deploy it for specific tasks. Lead author Nathan Ng, a graduate student at the University of Toronto and visiting student at MIT, emphasizes the importance of ensuring that models’ uncertainty aligns with human intuition, especially since models can perform well in some scenarios and fail in others.
Ng collaborated with Roger Grosse, an assistant professor at the University of Toronto, and Marzyeh Ghassemi, an associate professor at MIT, to produce this research, which will be presented at the International Conference on Machine Learning.
Uncertainty quantification methods traditionally involve complex statistical calculations that don’t scale well with large models. These methods also require assumptions about the model and its training data, which can limit accuracy. The MIT team adopted a different strategy, employing the minimum description length principle (MDL). MDL sidesteps the need for these assumptions, enhancing the quantification and calibration of uncertainty for test points the model labels.
The researchers developed a technique called IF-COMP, which makes MDL feasible for large deep-learning models used in real-world applications. MDL evaluates all possible labels a model could assign to a test point. If many labels fit well, the model’s confidence in its chosen label should decrease. Essentially, a model’s confidence can be gauged by its willingness to update its belief when given counterfactual information.
In practice, if a model labels a medical image as showing a pleural effusion but can be easily convinced it shows an edema instead, its initial confidence should be low. Using MDL, a confident model uses a short code to describe a data point, while an uncertain model uses a longer code, reflecting multiple possible labels. This code length, or stochastic data complexity, decreases if the model is confident when given contradictory evidence.
Testing each data point with MDL traditionally demands immense computation. However, IF-COMP introduces an approximation method using an influence function, combined with a technique called temperature-scaling, to improve model calibration. This blend allows for accurate, efficient estimations of stochastic data complexity.
IF-COMP can thus produce well-calibrated uncertainty quantifications efficiently, detect mislabeled data points, and identify outliers. The researchers tested IF-COMP on these tasks, finding it faster and more accurate than existing methods.
Marzyeh Ghassemi highlights the growing need for tools that audit and ensure model calibration, especially as machine-learning models increasingly impact human-centric decisions. IF-COMP’s model-agnostic nature means it can provide accurate uncertainty estimates across various machine-learning models, broadening its real-world applicability.
Ng cautions that while models can appear confident, they might still entertain numerous contradictory possibilities. Future research will explore applying this approach to large language models and other potential uses for the minimum description length principle.
