Y. Qin "A Path to Trustworthy Foundation Models"

ECE Assistant Professor Yao Qin’s group addresses degradation in performance stemming from robustness issues to help treat those with type-1 diabetes

photo of qin and her research group

From The ECE Current (2023-24) Newsletter

Deep Neural Networks have achieved significant success, yet they continue to suffer from various robustness issues, especially when being deployed in the real world. For example, neural networks suffer from sensitivity to distributional shift, when a model is tested on a data distribution different from what it was trained on. Such a shift is frequently encountered in practical deployments and can lead to a substantial degradation in performance. In addition, models are often over-confident in their predictions. A well-calibrated model should have confidence aligned with its accuracy, but standard training methods do not inherently calibrate confidence in accordance with performance.

Furthermore, neural networks are vulnerable to adversarial examples – small perturbations to the input can successfully fool classifiers into making incorrect predictions. The susceptibility to adversarial attacks poses a great security risk for deploying ML systems, as these types of attacks can bypass human scrutiny. Taken together, there are significant challenges in building trustworthy foundation models for real-world deployment, especially in safety-critical applications. In order to achieve trustworthy foundation models, we aim to enhance robustness through every stage of model development, from safe data collection & augmentation, to efficient & robust training techniques, as well as developing comprehensive evaluations of model robustness. In addition, as robustness takes on paramount significance for safety-critical applications, our lab aims to unlock the power of data-driven machine learning for diabetes care. Type1 diabetes, in particular, poses unique challenges and demands precise management to ensure the well-being of patients. One crucial aspect of this management is insulin recommendation, a process that requires accuracy and safety to prevent potentially life-threatening complications. To address this challenge, we are embarking on a journey to develop machine learning models dedicated to insulin recommendation.

These models will automate the insulin recommendation process, reducing the burden on both patients and healthcare providers. In addition, we will prioritize safety and robustness in model design to ensure that the recommended insulin doses are aligned with patients’ specific needs and medical history. This will lead to more timely and efficient insulin adjustments, potentially improving the overall quality of life for individuals with type-1 diabetes.

The ECE Current (2023-2024) "A Path to Trustworthy Foundation Models” (page 6)