Skip to main content

Training Safe Machine Learning Models Using Mathematical Optimization

Primary supervisor

Buser Say

Co-supervisors

Research area

Optimisation

Machine learning models have significantly improved the ability of autonomous systems to solve challenging tasks, such as image recognition, speech recognition and natural language processing. The rapid deployment of such models in safety critical systems resulted in an increased interest in the development of machine learning models that are robust and interpretable

In the context of machine learning, robustness refers to the performance stability of the model in the presence of natural and/or adversarial changes. For example, random noise in the training dataset, or random changes in the environment can both significantly degrade the testing accuracy of a machine learning model when they are not accounted for. Orthogonal to robustness, interpretability is concerned with the insights that the learned model can provide to its users about the relationships that are present in the dataset or the model. Interpretable models are desirable as they often drive actions and further discoveries. This project will focus on training robust and interpretable machine learning models using mathematical optimization.

 Visualization of interpretability results for MNIST (from paper [1]) over each number 1, . . . , 9. A pixel is colored to: black if the learned weight has a value 1, white if the learned weight has a value -1, and gray otherwise. Intuitively, each subfigure visualizes what the learned model predicts each number to look like (and not to look like) based on its learned weights.
Visualization of interpretability results for MNIST (taken from the paper [1]) over each number {0, . . ., 9}. A pixel is colored to: black if the learned weight has a value 1, white if the learned weight has a value -1, and gray otherwise. Intuitively, each figure visualizes what the learned model predicts each number to look like (and not to look like) based on its learned weights.

 

URLs/references

[1] Sanjana Tule, Nhi Ha Lan Le, and Buser Say. Training Experimentally Robust and Interpretable Binarized Regression Models Using Mixed-Integer Programming. In arXiv, 2021.

Required knowledge

A successful candidate should have strong programming skills (i.e., in Python) as well as experience in the following areas:

  • training machine learning models, and
  • mathematical optimization.

Project funding

Other

Learn more about minimum entry requirements.