Humans have the remarkable ability to continually learn and adapt to new scenarios over the duration of their lifetime. The rapidly growing field of lifelong learning in AI aims to mimic this very ability through constant and incremental development of increasing complex behaviors after deployment in the real-world. Lifelong learning systems continually consolidate new information to improve performance while minimizing loss of past learning, transfer knowledge from learned tasks to new tasks, and are sustainable over an arbitrarily long life with limited resources (e.g., memory, energy, time).
The conditions enforced by lifelong learning are the following:
Continual learning: They learn from a non-stationary stream of data (both novel and recurring) by consolidating new information while minimizing catastrophic forgetting of prior knowledge without distinct training and testing phases.
Transfer and adaptation: They perform better on an average on both novel and known tasks and maintain performance during rapid changes in an ongoing task (adaptation).
Sustainability: They can learn for an arbitrary long lifetime using limited resources (e.g. memory, time, energy) in a scalable manner.
Lifelong learning faces the well-known constraint for both biological and artificial neural networks called the stability-plasticity dilemma. Learning requires plasticity for integration of new knowledge while also stability to prevent forgetting of old knowledge. However, an extreme condition called catastrophic forgetting occurs when learning a new task using deep neural networks in contrast to biological systems where forgetting is typically gradual.
In this master's thesis, the student will investigate a method called elastic weight consolidation in the SINTEF machine learning pipeline Erdre https://github.com/SINTEF-9012 to realize lifelong learning. The basic idea behind elastic weight consolidation is to lower the learning rate of some weights while increasing the learning rate of others to adapt to new data. The pipeline will be tested on various case studies such as manufacturing, occupational health, Internet of Underwater Things stemming from various European Union and Norwegian research council projects.