Zero-Shot Learning

Dr. Timothy Hospedales
UK, School of Electronic Engineering & Computer Science, Queen Mary University of London
The classic paradigm of predictive modelling in supervised machine learning involves training classifiers or regressors to predict a target variable of interest based on large quantities of annotated training data. However, despite the age of “big data”, it is often the case that the specific category of interest to be recognised has few or no prior examples — as in the case of rare or recently emerged phenomena. This talk will introduce the new and exciting area of zero-shot machine learning, which addresses this setting of supervised prediction with zero prior training examples.
I will show how zero-shot learning can be achieved with a few strategies including via semantic attributes and distributed (vector space) models of words. We will see how zero-shot learning can be understood from a variety of perspectives including as an extreme form of classifier or regressor generalisation, learning a cross-modal embedding, or as a particular category of neural network.
Throughout the talk, I will introduce a variety of contemporary example applications of zero-shot learning including computer vision, forensics, and natural language processing. Finally, I will outline a variety of current research issues and open questions in zero-shot learning including unification of attribute and vector space approaches, transductive learning, and zero-shot domain adaptation.