My name is Liu Shikun (刘诗昆 in Chinese characters), currently a MRes in Advanced Computing student at Imperial College London. I am now affiliated with Dyson Robotics Lab and working with Prof. Andrew Davison. Before that, I graduated with Bachelor degrees in mathematics (with Honors) and electrical engineering at The Penn State University.
My long-term goal is to understand and recreate the visual world by finding the universal representations via machine learning methods. I divide this quest in the following research topics which I am interested in and currently working on.
Multi-task Visual Understanding
Multi-task learning is about performing well in various tasks by finding shared representations. Understanding how to share and how to deal with task imbalance may help us to re-introspect and understand generalisation. Multi-task learning is also highly related to transfer learning and continual learning which deals with problems in knowledge distillation and catastrophic forgetting.
Learning to Learn
Learning to learn i.e. meta-learning and other meta-processes, such as 'learning to plan' and 'learning to communicate' is to design the model which recursively improves by learning onto itself. The fact that human-designed learning algorithms are usually not optimal encourages to the sense of evolutionary computing. Such methods include neural architecture search and model parameters optimisation from a hyper-network.