In traditional machine learning, a learning system can be trained to deal with a specific single task, while human is able to complete multiple tasks with the same learning strategy. To overcome the limitations in traditional machine learning, multi-task learning techniques are demanded by for artificial general intelligence (AGI). For AGI, a learning system works for various tasks by sharing relevant knowledge between tasks so that learning a new task is done more efficiently and the learning system can generalise better on multiple tasks.
This project is going to develop novel learning systems and learning algorithms for multi-task learning in terms of different learning paradigms ranging from supervised, unsupervised to reinforcement learning and their applications to real world problems. The main research theme is how to share the generic knowledge and the representations applicable to different tasks without scarifying the previous learning outcome for a specific task. In a lifelong learning setting that a new task is learnt by a system already trained on other tasks, harmless knowledge transfer is also an unsolved issue and hard to carry out in the use of deep learning for multi-task learning due to catastrophic interference. Furthermore, this project also needs to address common issues in machine learning such as domain shift. Regarding applications, multi-model information processing, robotics and general video game playing are among the proper test beds for different learning paradigms. 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.