A Continual Machine Learning Framework for Accelerating Scientific Discovery

Machine learning models, which have found tremendous success in several commercial applications where large-scale data is available (e.g., computer vision and natural language processing), are beginning to play an important role in scientific disciplines such as biomedicine. Over the past few years, several domain-specific knowledge discovery frameworks have been proposed. Despite significant advances made, current research trends in machine learning-based approaches have not kept pace, for two main reasons. 1) The rapid proliferation of biomedical literature (on average around 3,000 articles are published every day) necessitates the development of innovative systems that can continually acquire and adapt to the new data. However, the existing approaches usually adopt a static learning paradigm and thus are unable to handle this setting. 2) Since the existing approaches mainly assume a static setting, they do not factor in the temporal evolution of biomedical concepts. This is limiting because the biomedical concepts are known to periodically acquire new semantic sense and lose old ones. To address these aforementioned challenges, we propose to shift the research direction from the currently dominant paradigm of static learning to continual learning, wherein the proposed approach is able to transfer useful knowledge over time and process the newly available articles in an efficient yet accurate manner.
Specifically, the proposed approach exploits the unique capabilities of self-supervised learning, supervised learning and lifelong learning to design a continual learning framework that progressively acquires new scientific knowledge, models the semantic evolution of biomedical concepts, and generates actionable insights (novel meaningful associations) that can drive new research frontiers.