The Tradeoffs and Layered Architecture in Brain 100


Quanying Liu, Southern University of Science and Technology


2020.04.08 14:00-15:00



Zoom Info

Meeting ID: 296-479-454
Password: 024763


Nervous systems sense, communicate, compute, and actuate movement, using distributed components with tradeoffs in speed, accuracy, cost, sparsity, noise, and saturation throughout. Nevertheless, the resulting control can achieve remarkably fast, accurate, robust performance. We hypothesize it is due to a highly effective network and layered architecture that combines higher layers of planning/predicting with lower layer reflex/reaction. We proposed a theoretical framework using feedback control theory and information theory which connects the component level speed-accuracy tradeoffs (SATs) in neurophysiology and system level SATs in sensorimotor control performance. It provides a holistic perspective of both levels and is needed to clarify the properties of effective architectures, and why there is such extreme diversity across layers (from planning to reflex) and within levels (of sensorimotor systems and neural components). The results lead to a novel concept, ‘diversity-enabled sweet spots (DESSs)’: that is, an appropriate diversity in neurons/muscles across layers and within levels help achieve systems that are both fast and accurate despite being built from components that individually are not. DESSs explains the necessity of the observed nerve heterogeneity at the component level as well as the resultant performance heterogeneity at the system level.


Quanying Liu is an assistant professor at Department of Biomedical Engineering, Southern University of Science and Technology, PI of Neural Computing and Control Lab. Quanying had her bachelor in electrical engineering and master degree in computer science at Lanzhou University, China. She got her PhD degree at ETH Zurich, Switzerland in 2017, and then had the postdoc training at California Institute of Technology (Caltech), US. Her research focuses on computational modeling of brain, with two goals of better understanding the brain functions in computational terms and building more human-like intelligence in machines.