Project description

Zero-shot learning refers to a novel paradigm on learning how to recognise new concepts by just having a description of them. For example, zero-shot learning works on a setting of solving a classification problem when no labelled training examples are available for all classes, which are divided into two class subsets: training and unseen classes, where there are only examples of training classes available to be used in building up a classifier. Under the zero-shot paradigm, it is expected that a classifier trained on only the training-class examples works for test data of unseen classes by exploiting the side information regarding the semantic relationship between training and unseen classes. When this learning paradigm is used in multimedia information retrieval, there is a big challenge; i.e., there is a semantic gap between raw media data and their semantic meaning. As a new learning paradigm, zero-shot learning paves a new way to address issues such as a lack of training examples in supervised learning and expand the capacity of a learning system to deal with unknown situations as same as human beings do.

The project is going to investigate effective zero-shot learning techniques as fundamental research and their applications in real world problems such as multimedia information retrieval and multi-task reinforcement learning. The main research theme is how to bridge the semantic gap between raw media data descriptions and their semantic meaning. Surrounding this main theme, main issues to be studied include media representation learning, semantic representation learning and latent embedding frameworks with the state-of-the art deep learning and representation learning methodologies. As domain shift and the nature of high-dimensional data are generic issues in zero-shot learning, it is inevitable that this project has to address the common issues appearing in transfer learning and manifold learning. Regarding the applications to multimedia information retrieval, a real scenario based on video streams will be selected to be used as a test bed, which is a non-trivial part of this project. Likewise, zero-shot knowledge transfer in general video game playing could be a test bed for zero-shot reinforcement learning. It is worth mentioning that this project description is generic and a specific yet well-defined project needs to be developed based on a self-motivated student's own input.

In order to take this project, it is essential to have excellent mathematics and machine learning background knowledge as well as good programming skills. If you are interested in this project, please first visit my research student page: http://staff.cs.manchester.ac.uk/~kechen/ for the required materials and information prior to contacting me.