Thanks to advances in deep learning techniques, remarkable improvements have been achieved in many recognitions problems over the past several years, on a variety of problems such as image classification, voice recognition, object detection, etc. Typically, these models build on deep neural networks and large-scale annotated training data sets.
Although such supervised learning approaches are very effective in the development of limited-domain models, there are very striking shortcomings compared to the human cognitive skills. To this end, the human cognition is largely believed to benefit from a number of factors, including the abilities to use training examples more efficiently, to benefit from every-day observations and to accumulate life-long experiences.
The main topic of this project is the development of generative model based meta-learning approaches, towards enabling more data-efficient machine learning approaches.
The project has started in July 2020. More information about the project will be progressively added to this page.