Lifelong Learning of Discriminative Representations
March 25, 2014, 12:10 p.m. - March 25, 2014, 12:40 p.m.
We envision a machine learning service provider facing a continuous stream of problems with the same input domain, but with output domains that may differ. Clients present the provider with problems implicitly, by labeling a few example inputs, and then ask the provider to train models which reasonably extend their labelings to novel inputs. The provider wants to avoid constraining its users to a set of common labels, so it does not assume any particular correspondence between labels for a new task and labels for previously encountered tasks. To perform well in this setting, the provider needs a representation of the input domain which, in expectation, permits effective models for new problems to be learned efficiently from a small number of examples. While this bears a resemblance to settings considered in previous work on multitask and lifelong learning, our non-assumption of inter-task label correspondence leads to a novel algorithm: Lifelong Learner of Discriminative Representations (LLDR), which explicitly minimizes a proxy for the intra-task small-sample generalization error. We examine the relative benefits of our approach on a diverse set of real-world datasets in three significant scenarios: representation learning, multitask learning and lifelong learning.