Continuous Learning at the Edge/IoT devices using Meta Learning

Continual Learning is an area of machine learning which deals with learning through time, with changes in environment and tasks to learn. For example, a typical continual learning problem would be to train a sensor to adapt its behaviour to changes of weather and location. The goal of continual learning is through learning a sequence of tasks and/or environments be able both to apply the knowledge gained to perform better on future tasks (forward transfer) but also to remember and improve upon the performance of previously seen tasks which will be needed in the future (backwards transfer). Due to the broad scope of continual leaning, we restrict the focus of the project to look at continual learning methods which help solve the real-world problems that arise in edge devices and sensors. In this scope many interesting challenges appear, which includes but is not limited to resource constrained devices, domain shift and limited supervision.

The aim of this project is to explore and design methods for solving continual learning problems which more directly address the needs of real-world edge devices than previous work. We will achieve this aim by creating realistic settings for continual learning, which are inspired and applicable to problems appearing in edge systems, and then aim to create high performing methods for solving such settings. Our settings may contain non-trivial transfer between tasks, periodic occurrences of tasks, memory limited agents, limited data, weak supervision and potentially many other factors necessary for edge device applications. Creating these settings will then enable us to create and evaluate methods that are better suited to tackle the challenges faced in the real world, particularly those faced by edge devices.

We propose to evaluate our work using the standard approach in continual learning and experimental machine learning in general. This comprises mainly of evaluating our methods across several datasets/benchmarks and against numerous other methods, with the potential of providing theoretical results when necessary and possible to back up the empirical evidence. Due to the fact that as part of the project we propose creating new settings for continual learning it is possible that we will have to create some benchmarks to be able to evaluate the performance of models in these settings. Furthermore, we will probably also need to create new evaluation criteria to accurately reflect performance in these new scenarios. Therefore, a key part of the project may be to design appropriate benchmarks and evaluation criteria.