In this proposal the candidate introduces the idea to perform his work in a hot research topic in the field of Machine Learning and Robotics such as Continual Learning. The broad aim of the project is to let autonomous robots be able to interact with highly dimensional environments such as the real world in an incremental fashion, minimizing the amount of human supervision as much as possible. An efficient acquisition and storage of information is the first critical step to give robots the capability of building-up their own experience. The long-standing challenge that prevents the construction of an incremental knowledge base is known in the literature as catastrophic forgetting. Indeed, when tasks are presented to the Machine Learning model in a sequential way, new knowledge interferes with the stored one, leading to a fast performance degradation for the older functions.
The second critical step to ensure Continual Learning regards the ability to exploit past information to speed up learning in a related task. When moving from a domain to another, a knowledge transfer is needed not only to reduce the amount of computations performed by the algorithm but also to decrease human effort in supervised scenarios, where extensive data labeling is generally needed. The efficient application of transfer learning would benefit the robotic domain. In fact, two classical robotic tasks such as navigation and control are extremely time consuming if the algorithm must be retrained every time.
With the previously mentioned objectives the candidate is proposed to positively contribute to the field of Continual Learning. Trivially, this is an arduous research problem that is still far from being solved. Big projects are trying to play a role in different aspects of the field making slow but valuable contributions. However, some key aspects are expected to bring back constructive notions that would benefit the entire field. First, a clearer picture of Deep Learning mechanisms would help to understand the limitations of current Machine Learning models. Second, an extensive study of the literature along with the addition of novel discovered concepts is expected to help the design of frameworks that outperforms the old ones. These novel frameworks are foreseen to:
- classify/predict new instances of already learned classes/functions, implying good generalization capabilities
- classify/predict patterns of new classes/functions without losing performance on previous once
- classify/predict patterns of new classes/functions at a faster rate
- allow a real robot to perform the tasks written in previous items
Finally, the publication of datasets regarding less popular but still crucial domains of Continual Learning would enhance the scope of the research enabling worldwide scientist to test their algorithms on a common benchmark.
To give a clearer picture of the possible impacts and innovations of the expected results, a complete problem coming from the emerging field of Service Robotics is reported.
In this area, autonomous robots must be able to cooperate with humans in dynamic environments (home, workplace) providing useful notions or services to requesting users. As an example, a robot is responsible for bringing any kind of asked portable object to a user. For this reason, the service robot is expected to manage a visual, a localization, a control and a natural language module. Clearly, these modules interact among each other to perform high level actions. The first one is needed to perceive
the world, including the desired object and the interacting person. Continual learning will help the robot to recognize and recall already requested items. The second module is foreseen to allow robotic self-localization and object localization in wide indoor scenarios. The service robot is expected to use extracted structures or old memories to improve localization performances in unseen rooms or to recall the most probable requested article position. An incremental learning control module would help robots to handle diverse objects over time and to learn physical interactions. Lastly, language models for objects names and action commands are required to be updated over time. Only the interplay of multiple long-term processing units will allow autonomous robots to solve complex tasks in highly dimensional and unpredictable environments. These complex implementations will push forward the current state of the art that has been mainly focused over classification tasks in the static visual domain and much less over the dynamic problems that belong to the robotic domain.