Most machine learning systems are "one-shot" meaning that they learn a single concept from data that are presented all at once. However, an intelligent agent like a robot must collect examples of concepts one by one in a complex environment that may be changing. The learning system must have properties that most common ones do not. It must be "incremental", accumulating knowledge and building on already learnt concepts. It must be able to recover from mistakes and cope with a dynamic environment. It can also use "active learning". That is, the robot can conduct experiments to test hypotheses.
This project builds on previous work by students at UNSW. The aim is to equip one of our robots with some of the above capabilities.