Grant: Discovery Grants Program - Individual

Machine learning research has focused on inducing a model of a function or classification task strictly from a set of training examples. Consequently, machine learning systems do not take advantage of previously acquired task knowledge when learning a new and potentially related task. This is unfortunate since prior task knowledge offers the potential to produce more accurate models in less time and using fewer training examples. My research investigates machine life-long learning (ML3) systems (particularly, architectures of artificial neural networks) that are capable of retaining and selectively transferring task knowledge when learning a new task. Data mining and adaptive software systems can benefit from the use of inductive transfer and ML3 methods. For example, more accurate medical diagnostic models can be developed from small samples of patients by using previously learned and related diagnostic models. Adaptive systems can learn more efficiently the interaction model (profile) of a new user by selectively transferring knowledge from similar user models. This research program will explore several questions raised by my most recent research into the use of multiple task learning (MTL) methods: What characteristics must a machine learning method have to affect inductive transfer using contest sensitive MTL (csMTL) encoded examples? Under what conditions will the csMTL approach fail? What is the best use of available training examples? Does the csMTL method mitigate the need for a measure of task relatedness? Can the approach be used for domains with real-valued outputs and multiple outputs per task? Can csMTL be used to accurately consolidate task knowledge? The return on investment in this research will be advancements in machine learning theory, the creation of a valuable machine learning research system that will be shared with the community, the training of several high quality researchers and the application of next generation machine learning to data mining and adaptive software systems.