PIMS Lunchbox Lecture: Quan Long

Machine Learning including deep learning techniques have been successfully used in many big-data fields. However, a limitation of many machine learning tools is that one needs to have a very large sample size to train a model with many parameters. This may prevent the broader use of machine learning in sample-sparse domains. For instance, in medical genetics, the number of patients of a particular disease available for a research project may be at the level of hundreds or even dozens, which is way lower than the requirement of many machine learning techniques that are sample-hungry. Towards this line, researchers have developed a technique called “transfer-learning”, which can re-task an established general model (that are usually trained by very large sample) to a specific target using tailored samples of limited size. Such transfer-learning models open the door of developing many tools tailored to specific tasks using small samples with nimble training procedure. In this talk, I will first explain the basic theory of transfer-learning, followed by an introduction of its use in computer science including natural language process. I will also present a project using transfer-learning to characterize genetic basis of complex diseases by retasking a large general model.