Cognitively Inspired Machine Learning for Biomedical Sciences

Abstract:

Our team at National University of Singapore utilizes newest discovery from cognitive sciences especially consciousness related knowledge to inspire next generation of machine learning and apply the methods to understand biomedical sciences. In this talk we focus on using method inspired by the global workspace theory in cognitive neuroscience, we present CellMemory, a bottlenecked Transformer designed to comprehend cell representations for single-cell data integration, annotation and reference mapping. We discovered that CellMemory, even without pre-training, exhibited enhanced generalization capabilities compared to Large Language Models pre-trained with tens of millions of cells. By integrating a lung atlas, CellMemory provided insightful interpretations for malignant cells outside the reference and was able to display the internal reasoning patterns of CellMemory in an interpretable manner. Lastly, by integrating a population-level cell atlas of Alzheimer's disease, we confirmed that CellMemory can perform granular-level integration and annotation of single-cell and spatial data.