Invited Lecture I - August 3, 2012
We will cover machine learning fundamentals, recent work on transfer learning, and its importance in the larger contexts of Machine Lifelong Learning and AGI. More specifically, Context-sensitive Multiple Task Learning, or csMTL, is presented as a method of transfer learning which uses a single output neural network and additional contextual inputs for learning multiple tasks. Motivated by problems with the application of standard MTL networks to machine lifelong learning, csMTL was developed and found to improve predictive performance for a primary task from impoverished training sets when learning in the presence of related tasks. We will show that the reason for this performance improvement is a reduction in the number of effective free parameters in the csMTL network brought about by the shared output node and weight update constraints due to the context inputs. An examination of other ML models developed from csMTL encoded data provides initial evidence that this improvement is not shared across all machine learning algorithms. Connections to recent work in Deep Learning Architectures will be made. And an interesting csMTL application that transfers knowledge between image transformation tasks will be demonstrated using some familiar facial images.