Interpretable and Fast Prediction Models (Liang Lan et al.)

Due to recent advances in training scalable machine learning models (e.g. deep neural networks, kernel classifiers, random forests) on big data, machine learning algorithms have been applied to many domains and achieved excellent performance in a wide variety of real-life tasks. However, there are still two main obstacles that prevent successful deployment of machine learning models: (1) model interpretability; and (2) prediction time and space complexity. Interpretability is an essential requirement for deploying machine learning models in many domains, especially in biomedical diagnosis and criminal justice. In such domains, the decision makers will not use a machine model that they can not understand and trust. However, the increasing complexity of machine learning models is marking them harder to explain. On the other hand, low prediction time and space complexity is also a very critical requirement. Unlike the training stage that can take several hours (even several days), test stage usually requires real-time processing. The problem is more challenging when prediction models run on limited computing resources (e.g. mobile device)

The objective of this project is to overcome the above mentioned two obstacles of deploying machine learning models. We will develop novel model compression approaches to approximate complex prediction models using interpretable and simple models with lower time and space complexity. Specifically, we will use simple and interpretable models (e.g. rule-based models, tree-based models, linear models) to approximate complex and accurate prediction models (e.g. deep neural networks, kernel classifiers, random forests) without sacrifice too much accuracy.

Related Publications:


For further information on this research topic, please contact Dr. Liang Lan.